Exchange Rate Predictability Rossi

UNIVERSITY OF SOUTHERN CALIFORNIA
Center for Applied Financial Economics (CAFE)
Exchange Rate Predictability

Barbara Rossi
Electronic copy available at: https://ssrn.com/abstract=2316312

Exchange Rate Predictability
Barbara Rossi
July 14, 2013
Abstract
The main goal of this article is to provide an answer to the question: “Does any-
thing forecast exchange rates, and if so, which variables?”. It is well known that
exchange rate ‡uctuations are very di¢ cult to predict using economic models, and
that a random walk forecasts exchange rates better than any economic model (the
Meese and Rogo¤ puzzle). However, the recent literature has identi…ed a series of fun-
damentals/methodologies that claim to have resolved the puzzle. This article provides
a critical review of the recent literature on exchange rate forecasting and illustrates
the new methodologies and fundamentals that have been recently proposed in an up-
to-date, thorough empirical analysis. Overall, our analysis of the literature and the
data suggests that the answer to the question: "Are exchange rates predictable?" is,
"It depends" –on the choice of predictor, forecast horizon, sample period, model, and
forecast evaluation method. Predictability is most apparent when one or more of the
following hold: the predictors are Taylor rule or net foreign assets, the model is lin-
ear, and a small number of parameters are estimated. The toughest benchmark is the
random walk without drift.
Keywords: Exchange Rates, Forecasting, Instability, Forecast Evaluation
Acknowledgments: I thank D. Ferraro for assistance in collecting the data used
in this paper, Janet Currie and four anonymous referees for their generous and useful
comments, and Greg Ganics for assistance in proofreading.
J.E.L. Codes: F3, C5
Barbara Rossi (ICREA-UPF, Barcelona GSE, CREI). Mailing address: Universitat Pompeu Fabra-
CREI, Carrer Ramon Trias Fargas, 25-27, Mercè Rodoreda bldg., 08005 Barcelona SPAIN. Tel.: +34-93-
542-1655; e-mail: barbara.rossi@upf.edu

1 Introduction
The objective of this article is to o¤er a critical survey of the literature on predicting exchange
rates in the last ten years. Since Meese and Rogo¤ (1983a,b, 1988), it has been well known
that exchange rates are very di¢ cult to predict using economic models; in particular, a
simple, a-theoretical model such as the random walk is frequently found to generate better
exchange rate forecasts than economic models. The latter …nding is known as "the Meese
and Rogo¤ puzzle". It is important to note that Meese and Rogo¤’s (1983a,b) …nding that
the random walk provides the best prediction of exchange rates should not be interpreted as
a validation of the e¢ cient market hypothesis. The e¢ cient market hypothesis states that, in
the absence of risk premia or when time variation in risk premia tends to be small relative to
variation in fundamental pricing factors, bilateral exchange rates are the market’s best guess
of the relative, fundamental value of two currencies based on all available information at that
time. The e¢ cient market hypothesis does not mean that exchange rates are unrelated to
economic fundamentals, nor that exchange rates should ‡uctuate randomly around their past
values. Hence the puzzle. However, the recent literature has identi…ed new macroeconomic
and …nancial predictors that claim to forecast exchange rates. The goal of this article is to
review both traditional as well as newly proposed exchange rate predictors and evaluate their
ability to forecast exchange rates. The main goal is to provide an answer to the questions:
“Are exchange rates predictable? And, if so, which predictors are the most useful to forecast
exchange rates?”.
When trying to answer these questions, a series of complications arise. First, a wide
variety of predictors, models, estimation methods, measures of predictive content as well as
evaluation tests have been used in the literature. Thus, researchers attempting to forecast
exchange rates need to make several choices, such as: Which predictors to use? Which
forecast horizon to predict? Which model to estimate? Which data frequency? Which
sample? One of the goals of this paper is to provide guidance to researchers on navigating
the existing literature as well as to provide a reliable overview of established …ndings that
can be helpful in making these choices. Second, existing papers rely on di¤erent predictors,
tests, samples or databases; it is possible that such predictors might have lost their forecasting
ability, or may not be robust to other databases or samples. In addition, while a predictor
might be successful according to a metric/test, it may not be so according to a di¤erent
one. We therefore perform a thorough empirical evaluation of the success of the predictors
identi…ed in the literature using the most recent techniques and databases. Thus, our article

starts with a critical overview of existing predictors and the empirical stylized facts identi…ed
in the literature, with particular emphasis on the last ten years. Then, we illustrate the
existing empirical evidence using the most up-to-date data and evaluation techniques in
order to answer the question: “Does anything forecast exchange rates?”.
This article should be of interest for several audiences. Economists and researchers in
academia will …nd that the literature review and the empirical investigation provide guidance
to navigate the literature, and will be useful for their work. Practitioners and forecasters
at Central Banks and private businesses will also be interested in knowing which predictors,
models and methodologies successfully predict exchange rates. Policymakers, for whom
successful policy decisions crucially depend on successful forecasts, should also be interested
in our assessment on where the literature stands. Finally, newspapers’frequent discussions
of exchange rate forecasting suggest this literature review would be useful beyond academia
and policy circles.
More in detail, why are exchange rate forecasts useful for Central Banks and policymak-
ers? Wieland and Wolters (2011) provide a detailed review on how forecasts are used in
policymaking. Typically, forecasts are used to project the consequences of particular pol-
icy measures for policymakers’ targets. According to Greenspan (1994, p. 241), "implicit
in any monetary policy action or inaction is an expectation of how the future will unfold,
that is, a forecast". Wieland and Wolters (2011) provide empirical evidence that Central
Bank policies in the US and Europe are described by interest rate rules, where interest rates
respond to forecasts of in‡ation and economic activity, rather than outcomes. Not only eco-
nomic policy relies on macroeconomic forecasts: the path of the policy may directly a¤ect
the forecasts ("projections") of macroeconomic aggregates. In the US, for example, prior
to each Federal Open Market Committee meeting,1 the Federal Reserve sta¤ produces fore-
casts of several macroeconomic aggregates at horizons up to two years as a basis for their
discussions. The variables forecasted by the sta¤ include exchange rates, which in‡uence
current account projections as well as US real GDP growth, eventually. As pointed out in
Edge et al. (2010), among others, the Federal Reserve sta¤ forecast is derived from data, a
variety of models and forecasting techniques, as well as expert judgment. Typically, these
forecasts are conditioned on a speci…c future time path for the federal funds rate, the main
instrument of monetary policy. The policy scenarios considered by the Federal Reserve sta¤
may also include dollar depreciation/appreciation scenarios (i.e. scenarios in which the dollar
appreciates or depreciates more than in the baseline forecasts, where typically it is assumed
1
The Federal Open Market Committee is the meeting where US monetary policy is decided.

constant, according to the random walk model). Policy decisions are then taken on the basis
of what the policymaker deems the most likely scenario. Exchange rate projections are also
especially important for Central Banks of countries that are heavy importers/exporters of
commodities. For example, one of the models used at the Bank of Canada is Amano and
van Norden’s (1995), where real exchange rates depend on terms of trade (see Coletti and
Murchison, 2002).
At the same time, it should be noted that this review has an empirical, "reduced-form"
focus. There are several reasons behind this choice. First, the majority of the empirical
work in this area is done with a reduced-form approach; second, while there are theoretical
structural models of exchange rate determination, typically they are too stylized to be liter-
ally taken to the data and successfully used for forecasting exchange rates. Moreover, fully
developed structural models typically do not …t exchange rate data well, not to mention
forecast them. Thus, while this article will sketch several "theoretical" models of exchange
rate determination, this discussion is mainly provided to motivate the choice of economic
predictors that have been considered in the literature. Furthermore, throughout the paper
we focus on monthly and quarterly frequencies, as they are the ones of interest to economists;
we will not consider very high frequency data analyses that are instead mostly of interest to
risk management and …nance. Finally, there is a large literature on in-sample estimation of
exchange rate models. In-sample …t does not necessarily guarantee out-of-sample forecast
success, as we will discuss. Thus, in this overview we will mainly focus on out-of-sample
forecasts, although we will provide some discussion of in-sample …t. Note that, typically,
real exchange rates are …tted in-sample, while nominal ones are forecasted out-of-sample;
therefore, we will focus on the latter.2
Overall, our analysis of the literature and the data suggests that the answer to the
question: "Are exchange rate predictable?" is, "it depends". In fact, it depends on the
choice of predictor, forecast horizon, sample period, model, and forecast evaluation method.
Predictability is most apparent when one or more of the following hold: the predictors are
Taylor rule and net foreign assets fundamentals, the model is linear, and a small number
of parameters are estimated. The toughest benchmark is the random walk without drift.
There is some instability over samples for all models, and there is no systematic pattern
across models in terms of which horizons or which sample periods the models predict best.
Among the negative …ndings on which the literature has reached a consensus, typically, PPP
and monetary models have no success at short (less than 2-3 years) horizons.
2
See Rogo¤ (1996) for a review of in-sample …t of real exchange rate models.

More in detail, we draw …ve general conclusions.
First, the degree of success in forecasting exchange rates out-of-sample does depend
on the choice of the predictor. Although there is disagreement in the literature, overall
the empirical evidence is not favorable to traditional economic predictors (such as interest
rates, prices, output and money).3 Instead, Taylor-rule fundamentals and net foreign asset
positions have promising out-of-sample forecasting ability for exchange rates. The consensus
in the literature is that the latter fundamentals have more out-of-sample predictive content
than traditional fundamentals; the disagreement in the literature is in the degree to which
they can resolve the Meese and Rogo¤ puzzle.
Second, overall, among the model speci…cations considered in the literature, the most
successful are linear ones.4 Typically, in single-equation linear models, the predictor choice
matters more than the model speci…cation itself.5
Third, data transformations (such as de-trending, …ltering and seasonal adjustment) may
substantially a¤ect predictive ability, and may explain di¤erences in results across studies.6
Another important factor that, for some fundamentals, may a¤ect predictive ability is the
use of real-time rather than revised data.7 For a given model and predictor, predictive ability
seems also to depend on the choice of the country. On the other hand, the frequency of the
data does not seem to a¤ect predictability.
Fourth, empirical results vary with the benchmark model, the sample period, forecast
evaluation method and the forecast horizon. The random walk consistently provides the
toughest benchmark. Di¤erent models vary in terms of which sample periods and forecast
horizons work best, with no apparent overall pattern.
Finally, on the one hand, our empirical analysis con…rms several …ndings in the lit-
erature: while several predictors display in-sample predictive ability for future exchange
rates, only Taylor-rules display consistently signi…cant out-of-sample forecasting ability at
short horizons; and panel monetary models display some forecasting ability at long horizons.
Furthermore, our analysis reveals instabilities in the models’ forecasting performance: the
predictability of fundamentals varies not only across countries, models and predictors, but
3
Except possibly for monetary fundamentals at long horizons and interest rates at short horizons.
4
Among them, error correction models (either single-equation or panel) are successful at long horizons,
although there is disagreement among researchers regarding the degree of robustness of the result.
5
For example, whether the researcher uses contemporaneous, realized or lagged fundamentals.
6
For example, predictability of the monetary ECM model is much weaker or completely disappears after
estimating the cointegrating parameters.
7
This is a concern for monetary fundamentals but less of a concern for Taylor-rule fundamentals.

also over time. None of the predictors, models, or tests systematically …nd empirical support
of superior exchange rate forecasting ability across all countries and time periods: when pre-
dictability appears, typically it occasionally does so for some countries and for short periods
of time. These …ndings lead to new challenges: why does predictability change over time? Is
it possible to design ways to exploit instabilities to improve exchange rates’forecasts? We
discuss a few of these challenges in the last section.
There are several literature reviews on exchange rate predictability. How is this literature
overview di¤erent from the existing ones? Frankel and Rose (1995) review the empirical
literature on exchange rates up to 1995, whereas we focus on more recent contributions and
include a thorough empirical analysis which includes recent data as well as predictors that
have been identi…ed in the last decade. Engel, Mark and West (2007) focus on explaining
the ‡uctuations of exchange rates using selected models, countries and fundamentals; our
analysis considers a broader set of fundamentals and more recent data. Melvin, Prins and
Shand (2011) focus on forecasting exchange rates from an …nancial investor’s point of view,
e.g. carry trades; we focus instead on forecasting exchange rates using economic models and
macroeconomic predictors. Lewis (1995) focuses on reviewing several puzzles in international
…nancial markets, especially risk premia and home bias; we focus instead on assessing stylized
facts on exchange rate predictability. Finally, Rogo¤ (1996) and Froot and Rogo¤ (1995)
focus on purchasing power parity (PPP), according to which exchange rate ‡uctuations
re‡ect ‡uctuations of countries’relative prices; while we do consider PPP as a predictor, our
broader analysis includes several other predictors that have been considered in the literature
to forecast exchange rates.8
The article is organized as follows. Section 2 provides a simple guiding example. Sections
3 to 6 review the literature: Section 3 discusses the predictors that have been used to forecast
nominal exchange rates; Section 4 reviews a variety of models estimated in the literature;
Section 5 examines the characteristics of the data used in the literature; and Section 6
overviews the forecast evaluation methods. At the end of each section, we provide a summary
of the main …ndings in the literature. Finally, Section 7 revisits the empirical evidence in an
up-to-date, thorough empirical exercise.
8
For shorter reviews on exchange rate predictability that focus on speci…c topics, see: Neely (1997) on
technical trading rules; Bailliu and King (2005) on exchange rate models and the Canadian experience; Neely
and Dey (2010) on the e¤ects of macroeconomic news announcements on exchange rates; Chinn (2011) for
a survey on macroeconomic models of exchange rate determination; and Neely and Sarno (2002) on the
empirical performance of the monetary model.

2 A Guiding Example
Before reviewing in detail the predictors, models, data and forecast evaluation methods
proposed in the literature, let us consider a simple example to …x ideas about several basic
concepts that we will discuss in this review. Readers that are familiar with basic in-sample
…t and out-of-sample forecast methodologies can move directly to the next section.
Let the (log) of the exchange rate be denoted by st and let the (univariate) predictor (or
fundamental) be denoted by ft . Examples of predictors used in the literature are discussed
in detail in the next section, and the choice of the data is discussed in Section 5.
The relationship between the exchange rate and its fundamental can be described by
several models (see Section 4). For expositional purposes, let the model be linear and such
that it does not include a constant term:
Et (st+h st ) = ft ; t = 1; 2; :::; T;
where T is the total size of the available sample and h is the forecast horizon.
The model’s performance is typically evaluated relative to that of a benchmark model.
Let the benchmark model be the random walk without drift:
Et (st+h st ) = 0:
In fact, we will argue in Section 6 that the random walk without drift is the appropriate
benchmark for the analysis.
The predictive ability of the fundamental can be evaluated according to in-sample …t or
out-of-sample forecast performance. In-sample …t is typically evaluated by estimating over
the full sample,
! 1 !
X
T X
T
bT = ft2 ft (st+h st ) ;
t=1 t=1
and calculating a t-test on : if the fundamental contains relevant information, then should
be di¤erent from zero. The latter is known as an in-sample (traditional) Granger-causality
test. If the test rejects, it signals that the predictor contains useful information for explaining
exchange rate ‡uctuations over the full sample. However, this does not necessarily mean that
the predictor contains useful information to predict exchange rate ‡uctuations in real-time.
To assess the latter, it is common to turn to forecasting.
To evaluate the models’ out-of-sample forecasting ability, the sample is split into two
parts: the in-sample portion, consisting of observations from 1 to R, and the out-of-sample

portion, of observations R+h to T +h, of size P T R+1. In the rolling window forecasting
scheme, the parameter is re-estimated over time using the most recent R observations, where
R is known as the estimation window size:
! 1 !
X t X t
bt = f 2
fj h (sj sj h) ; t = R; R + 1; :::; T;
j h
j=t R+h+1 j=t R+h+1
to obtain a sequence of P h-step-ahead out-of-sample forecast errors, "ft+hjt st+h st bt ft ,

t = R; R + 1; :::; T . Note that the random walk forecast error is simply "rw
t+hjt st+h st .
Under the rolling window forecast scheme, the model parameters are re-estimated pro-
gressively over time. An alternative forecast scheme is the recursive one, where the model
parameters are always re-estimated using all the previous observations. That is, bt =
Pt 2
1 P
t
j=h+1 fj h j=h+1 fj h (sj sj h ) ; t = R; R + 1; :::; T:
The forecasting ability of the model is measured by a loss function; for example, a common
choice is the Root Mean Squared Forecast Error (RMSFE), which will be the objective of
our analysis unless otherwise noted:9
1 X f
T
2
RM SF Ef " :
P t=R t+hjt
1
PT 2
The model forecasts better than the random walk if RM SF Ef < RM SF Erw P t=R "rw
t+hjt :
To judge whether the model forecasts signi…cantly better, one typically tests whether RM SF Ef
RM SF Erw is equal to zero against the alternative that the di¤erence is negative, i.e. using
a t-test. Several methods to compute the standard errors and other available test statistics
are discussed in Section 6.
3 Which Predictors To Use?

This section reviews several economic predictors of exchange rates that have been used in
the literature. It explains why, according to economic theory, they should forecast exchange
rates, reviews papers proposing/using them, and summarizes their empirical …ndings.10 A
convenient summary of the various predictors used in the literature is provided in Tables 1
and 2. Some empirical …ndings are contradictory: the next sections investigate in detail the
reasons behind the di¤erences among the papers.
9 1
PT f
Other loss functions are sometimes used, such as the Mean Absolute Error, M AEf P t=R "t+hjt .
10
For a more detailed introduction to these models, see Obstfeld and Rogo¤ (1996) and Mark (2001).

3.1 Traditional Predictors
The traditional predictors used in the literature include interest rates, prices, money and
output di¤erentials.
(a) Interest Rate Di¤erentials
Uncovered interest rate parity (UIRP) dates back to Fisher (1896) –see Dimand (1999).
Fisher (1896) provided a general analysis of how interest rates can be related to expected
changes in the relative value of units of account or commodities: one of the examples he
considered concerned international currencies, and has become known as UIRP. UIRP states
that, in a world of perfect foresight11 with a nominal bilateral exchange rate St , investors
can buy 1=St units of foreign bonds using one unit of the home currency, where St denotes
the price of foreign currency in terms of home currency. Let the foreign bond pay one unit
plus the foreign interest rate between time t and t + h, it+h . At the end of the period, the
foreign return can be converted back in home currency and equals St+h 1 + it+h =St in
expectation. By arbitrage and in the absence of transaction costs, this return must be in
expectation equal to the return of the home bond, 1+it+h . That is, 1 + it+h Et (St+h =St ) =
1 + it+h , where Et (:) denotes the expectation at time t. Finally, by taking logarithms and
ignoring Jensen’s inequality, the previous UIRP equation can be rewritten as:
Et (st+h st ) = + it+h it+h ; (1)
where st ln (St ), = 0 and = 1; and h is the horizon. Similarly, covered interest rate
parity (CIRP) predicts exchange rates according to: Et (st+h st ) = + (Ft st ), where
Ft denotes the h-period ahead forward rate at time t.12
The empirical evidence is not favorable to UIRP. Meese and Rogo¤ (1988) use eq. (1) to
forecast real exchange rates out-of-sample using real interest rate di¤erentials, and compare
its performance with the random walk, …nding that the latter forecasts better. Similarly,
Cheung, Chinn and Pascual (2005) and Alquist and Chinn (2008) …nd that, although for
some countries UIRP forecasts better than the random walk at long horizons, its performance
is never signi…cantly better. Slightly more positive …ndings have been reported by Clark and
West (2006) at short-horizons, and Molodtsova and Papell (2009) for some countries.13 In-
sample empirical evidence is not favorable to UIRP either. The consensus is that, typically,
11
See Lewis (1995) for a discussion of UIRP without perfect foresight.
12
CIRP states that the spread between forward and spot exchange rates equals the nominal interest
di¤erential between two countries, and was developed by Keynes (1923).
13
Molodtsova and Papell (2009) estimate UIRP with unrestricted coe¢ cients (both constant and slope) as
well as without a constant and with an estimated slope. In the latter case, they only …nd marginal evidence

estimates of (1) display a negative and signi…cant slope, and a constant signi…cantly di¤erent
from zero; see Froot and Thaler (1990) for a survey.14
(b) Price and In‡ation Di¤erentials
According to Purchasing Power Parity (PPP), the real price of comparable commodity
baskets in two countries should be the same. That is, the price level in the home country,
converted to the currency of the foreign country by the nominal exchange rate, should equal
the price level of the foreign country. Thus, a unit of currency in the “home” country will
have the same purchasing power in the “foreign”country. The theory, due to Cassel (1918),
was developed during the debate on the appropriate value of nominal exchange rates among
major industrialized countries after the hyper-in‡ations in World War I. Let the logarithm
of the commodity price (CP) index in the home and foreign countries be denoted by pt and
pt , respectively. Then, PPP implies that:
st = + (pt p t ) + "t : (2)
where = 0 and = 1:
The out-of-sample empirical evidence is not favorable to PPP either: Cheung, Chinn and
Pascual (2005) …nd that, although PPP forecasts better than the random walk at the longest
horizons, its performance is never signi…cantly better; at shorter horizons, it is signi…cantly
worse than the random walk. Whether PPP holds in-sample is also debated. In particular,
two stylized facts emerge from Rogo¤ (1996). First, nominal exchange rates tend toward
purchasing power parity in the long run, although the speed of convergence is remarkably
slow. Second, short run deviations from PPP are substantially large. As Rogo¤ (1996) notes,
deviations from PPP can be attributed to transitory disturbances in the presence of nominal
price stickiness; thus, they should be short-lived (i.e., 1–2 years), while, in the data, half-
life deviations from PPP range between three to …ve years.15 Rogo¤ (1996, p. 647) called
this empirical inconsistency the “PPP puzzle.” Possible concerns and explanations include:
underestimation of the uncertainty around point estimates (Cheung and Lai, 2000, Kilian
of predictive ability for Australia and Canada; in the former case, they …nd strong evidence in favor of Japan
and Switzerland, as well as marginal evidence in favor of Australia and Canada.
14
Possible explanations include: the presence of a time-varying risk premium (Fama, 1984; Backus, Foresi
and Telmer, 2001); estimation biases (Bekaert and Hodrick, 2001); imprecise standard errors (Baillie and
Bollerslev, 2000, Rossi, 2007a); and small samples (Chinn and Meredith, 2004, who …nd positive evidence in
longer samples, and Chen and Tsang, 2011, who pool information from the whole term structure).
15
The half-life measures how many time periods it takes for the e¤ects of a shock to PPP to decrease by
50%.
10

and Zha, 2002, Murray and Papell, 2002, Gospodinov 2004, Lopez, Murray, and Papell,
2005, and Rossi, 2005c)16 and heterogeneity in disaggregate data (Imbs, Mumtaz, Ravn and
Rey, 2005).17
(c) Money and Output Di¤erentials
According to the monetary model of exchange rate determination, bilateral nominal
exchange rate ‡uctuations should re‡ect movements in countries’ relative money, output,
interest rates and prices. The monetary model was …rst introduced by Frenkel (1976) and
Mussa (1976), and builds on a simple small open economy model where real output is
exogenous. Real money demand is viewed as a function of income and the interest rate; by
using UIRP and PPP to substitute relative interest rates and prices as function of exchange
rates, one obtains a relationship between exchange rates, money and output di¤erentials.
More in detail, let mt be the logarithm of nominal money, yt be the logarithm of real output,
and the horizon, h, equals 1. Then, real demand for money is modeled as:
mt pt = it+1 + yt :
Assuming that a similar equation holds for the foreign country (mt pt = it+1 + yt ,
where for simplicity of notation we assumed that the coe¢ cients are symmetric and asterisks
denote foreign country variables) and taking the di¤erence between the two gives the relative
money demand equation: mt mt (pt pt ) = it+1 it+1 + (yt yt ). One approach
(valid if prices and exchange rates are completely ‡exible) is to assume that PPP holds at
every point in time, and substitute it in the relative money demand equation to get the
"‡exible price version of the monetary model":18
st = it+1 it+1 (yt yt ) + (mt mt ) : (3)
In the presence of sticky price adjustment, either the relative price level or in‡ation di¤eren-
tials are included as regressors to obtain the "sticky price version of the monetary model":
st = it+1 it+1 (yt yt ) + (mt mt ) + 2 (pt pt ) ; (4)

16
On the one hand, higher uncertainty implies that the PPP puzzle is even larger than previously thought;
on the other hand, this partially reconciles PPP with models with sticky prices, since con…dence intervals
also include very short half-lives.
17
Imbs, Mumtaz, Ravn and Rey (2005) …nd that, when heterogeneity is taken into account, the estimated
persistence of real exchange rates and the half-life falls dramatically. Using panel techniques, Taylor and
Sarno (1998) …nd evidence in favor of PPP in the long-run.
18
This was also referred to as the Frenkel-Bilson model; the coe¢ cient on money di¤erential is unity, due
to …rst degree homogeneity of relative money supply. See Meese and Rogo¤ (1983a).
11

in this case, PPP holds in the long run but does not hold in the short run.19
The empirical evidence on the monetary model is mixed. The in-sample evidence is
somewhat positive,20 while the out-of-sample evidence is less positive. On the one hand,
Meese and Rogo¤ (1983a,b) demonstrate that the random walk forecasts exchange rates
out-of-sample better than both the monetary models above, eqs. (3) and (4). Their …nding
has been con…rmed by Chinn and Meese (1995) for short horizon (one-month to one-year-
ahead) forecasts,21 by Cheung, Chinn and Pascual (2005), who …nd that the monetary model
does not predict well even at longer horizons (i.e. …ve years), and by Alquist and Chinn
(2008).22 Molodtsova and Papell (2009) also …nd very limited empirical evidence in favor
of the model.23 On the other hand, Mark (1995) …nds strong and statistically signi…cant
evidence in favor of the monetary model at very long horizons (i.e. three to four years).
The robustness of Mark’s (1995) …ndings has, however, been questioned by Berkowitz and
Giorgianni (2001), Kilian (1999), Groen (1999), Faust et al. (2003) and Rossi (2005a).
Overall, the former four papers …nd less evidence in favor of predictive ability even at long
horizons, whereas the latter …nds more positive results. The next sections will shed light on
the causes of the disagreement across the empirical …ndings.
(d) Productivity Di¤erentials
More general monetary models that include additional predictors have been considered
by several authors. For example, Cheung, Chinn and Pascual (2005) consider a model where
PPP does not hold even in the long run; instead, relative prices (pt pt ) are expressed as a
function of productivity di¤erentials zt following Balassa (1964) and Samuelson (1964):
st = it+1 it+1 (yt yt ) + 1 (mt mt ) + 2 zt + "t : (5)
Note that in the case of (5) the nominal exchange rate may depend on real variables. In
some studies, the real price of non-tradeables is included instead of productivity di¤erentials.
19
Sometimes in‡ation di¤erentials are used instead of prices (Cheung, Chinn and Pascual, 2005) and
coe¢ cients left unrestricted.
20
MacDonald and Taylor (1993), Husted and MacDonald (1998), Groen (2000, 2002), Mark and Sul (2001)
…nd cointegration between exchange rates and monetary fundamentals, while Sarantis (1994) does not. Rossi
(2006) rejects that the coe¢ cients of the monetary model are both constant and equal to zero, suggesting
time-varying predictive ability.
21
At longer horizons, Chinn and Meese’s (1995) results are slightly more positive, although statistically
signi…cant only for the Yen/US dollar exchange rate among the …ve currencies they consider.
22
Chinn and Meese (1995) consider eq. (3); Alquist and Chinn (2008) consider the sticky price monetary
model (eq. 4) with di¤erentials of money, real GDP, interest and in‡ation rates.
23
They …nd evidence only for two countries among the twelve they consider.
12

Cheung, Chinn and Pascual (2005) measure productivity di¤erentials by labor productivity
indices (real GDP per employee). They …nd that the model with productivity di¤erentials
does not forecast better than the random walk.24
(e) Portfolio Balance
Traditional portfolio balance models (Frankel, 1982; Hooper and Morton, 1982) include
a measure of stock balances:
st = 0 + 1 (it it Et (st+1 st )) + bt bt ; (6)
where bt is the stock of home assets held by home and bt is the stock of foreign assets held by
home, and the unobservable term Et (st+1 st ) is approximated by zero. Several measures
of balances have been used in the literature as broad proxies: cumulated trade balance
di¤erentials, cumulated current account balance di¤erentials, and government debt. Meese
and Rogo¤ (1983a,b) …nd that, even after augmenting the monetary model by a measure of
trade balance di¤erentials, the model still does not forecast better than the random walk, a
…nding con…rmed by Cheung, Chinn and Pascual (2005).25
Summary of Empirical Findings
Overall, the empirical evidence based on the traditional predictors is not favorable to
the economic models. While the out-of-sample forecasting ability of the economic predictors
occasionally outperforms that of a random walk in some studies for some countries/time
periods, it de…nitely does not systematically do so. More importantly, with a few exceptions,
their predictive ability is not signi…cantly better than that of a random walk at short horizons.
The main exception is the work by Clark and West (2006) regarding the out-of-sample
predictive ability of UIRP; the next sections will investigate the reasons why their …nding
is di¤erent from the rest of the literature. At longer horizons, there is more evidence of
predictive ability in favor of the monetary model, although the …nding is contentious. At
the same time, some predictors (i.e. interest rate di¤erentials) show signi…cant in-sample …t,
although with coe¢ cient signs that are inconsistent with economic theory.
24
Wright (2008) also includes productivity di¤erentials among his predictors. See Section 4 for a discussion.
25
Cheung, Chinn and Pascual (2005) estimate a model where exchange rate ‡uctuations are functions of
the government debt relative to GDP, the real interest rate, the net foreign asset position, the (log of the)
terms of trade, the (log) price level di¤erential and the relative price of non-tradeables.
13

3.2 Taylor Rule Fundamentals
Engel and West (2005, 2006) and Molodtsova and Papell (2009) propose fundamentals based
on a Taylor rule for monetary policy (Taylor, 1993). Taylor (1993) formalizes the idea that
the monetary authority sets the real interest rate as a function of how in‡ation di¤ers from
its target level (the higher the in‡ation, the more contractionary monetary policy will be)
and also as a function of the output gap ytgap (if output is below its potential, monetary policy
will be more expansionary). Taylor (1993) originally proposed the following speci…cation:
rt+1 = t
y
+ ytgap + ry ; where t is the in‡ation rate, y
is the target rate of in‡ation,
ytgap is the output gap, rt+1 it+1 t+1 is the real interest rate (de…ned as the di¤erence
between the nominal interest rate, it , and the in‡ation rate, t ), and ry is the equilibrium
real interest rate. Clearly, if one considers two economies, both of which set interest rates
according to a Taylor rule, by UIRP their bilateral exchange rate will re‡ect their relative
interest rates, and thus, as a consequence, their output gaps and their in‡ation levels. This
basic idea is at the core of the Taylor-rule fundamental model of exchange rates. We will
discuss this model in detail following the approach of Molodtsova and Papell (2009).
Molotdsova and Papell (2009) amend the Taylor rule to take into account two empirical
facts. First, in an open economy setting, as the Central Bank attempts to maintain the
nominal exchange rate at its purchasing power parity level (Svensson, 2000), monetary policy
also depends on the real exchange rate, qt st pt + pt . Second, interest rate changes are
sluggish since Central Banks prefer to avoid over-achieving their target (as in Clarida, Gali
and Gertler, 1998). By adding these features to the original Taylor rule, they obtain:
it+1 = (1 )( + t + ytgap + qt ) + it + vt+1
for all countries, whereas for the US = 0, and vt+1 is the monetary policy shock. That is,
using asterisks to denote foreign country variables:
it+1 = (1 )( + t + ytgap + qt ) + it + vt+1

it+1 = (1 )( + t + ytgap ) + it + vt+1 :
By taking the di¤erence of the two equations, using UIRP and re-de…ning the coe¢ cients,
one obtains the speci…cation in Molodtsova and Papell (2009):
Et st+1 st = e + eqt + e t + e ytgap + e t + eytgap + it it ; (7)
which they refer to as the “asymmetric” Taylor rule. They also consider imposing the
coe¢ cient to be the same and excluding the real exchange rate, a speci…cation they refer
14

to as the “symmetric, homogeneous”Taylor rule, and typically lagged interest rates are not
included:
Et st+1 st = e + e ( t t) + e (ytgap ytgap ) : (8)
Both the in-sample and the out-of-sample empirical evidence are mostly favorable to
Taylor-rule fundamentals, although with exceptions. Regarding the in-sample evidence,
Chinn (2008) estimates the Taylor model in-sample and …nds that the coe¢ cient signs are
not consistent with theory, and that the choice of the gap measure is not innocuous.26
Regarding the out-of-sample evidence, Molodtsova and Papell (2009) show that eq. (8)
forecasts exchange rates out-of-sample signi…cantly better than the random walk for several
countries, although the performance depends on the exact speci…cation.27 Molodtsova et al.
(2010), Giacomini and Rossi (2010) and Inoue and Rossi (2012) also …nd strong empirical
evidence in favor of Taylor-rule fundamentals. On the other hand, Rogo¤ and Stavrakeva
(2008) …nd that the empirical evidence in favor of Taylor-rule fundamentals is not robust.
Taylor rules are generally deemed to be a good description of monetary policy in the past
three decades, but monetary policy may have changed during the recent 2007 …nancial crisis.
Molodtsova and Papell (2012) study exchange rate forecasting during the …nancial crisis by
including indicators of …nancial stress in the Taylor rule, such as the Libor-OIS/Euribor-
OIS di¤erential, the Bloomberg and OECD …nancial condition indices, and the TED spread
di¤erential. Adrian et al. (2011) use instead measures of liquidity such as funding liq-
uidity aggregates of US …nancial intermediaries measured by stocks of US dollar …nancial
commercial paper and overnight repos. Both the latter papers …nd positive evidence.28
3.3 External Imbalance Measures

Gourinchas and Rey (2007) argue that not only the current account, but the whole dynamic
process of net exports, foreign asset holdings and return on the portfolio of net foreign assets
are important predictors of exchange rates. When a country experiences a current account
26
Chinn (2008) argues that this model of exchange rate ‡uctuations builds on UIRP: if the latter does not
hold in the data, it is surprising that the former holds. Engel and West (2006) calibrate the model in-sample
for real exchange rates and …nd a correlation of 0.5 between …tted and realized real exchange rates.
27
E.g. which output gap measure is used and whether a constant is estimated. Across the various
speci…cations they consider, typically, exchange rates of four to seven countries out of twelve are signi…cantly
predictable.
28
E.g. Adrian et al. (2011) …nd positive evidence for almost all the advanced countries as well as half of
the emerging countries they consider.
15

imbalance, the traditional intertemporal approach to the current account suggests that the
country will need to run future trade surpluses to reduce this imbalance. Gourinchas and Rey
(2007) argue instead that part of the adjustment can take place through a wealth transfer
between that country and the rest of the world occurring via a depreciation of the value of its
currency. Thus, they propose "net foreign assets" (NXA) as a potential predictor for future
exchange rate ‡uctuations. NXA is the deviation from trend of a weighted combination of
gross assets, gross liabilities, gross exports and gross imports, and measures the approximate
percentage increase in exports necessary to restore external balance, that is, to restore the
long run equilibrium of net exports and net foreign asset ratios.
The empirical evidence is overall favorable to external imbalance measures. Gourinchas
and Rey (2007) and Della Corte, Sarno and Sestieri (2010) …nd that the net foreign asset
model can predict (e¤ective) exchange rates out-of-sample signi…cantly better than the ran-
dom walk at both long and short horizons.29 Alquist and Chinn (2008) …nd that in some
sub-sample the net foreign asset model forecasts (bilateral) exchange rates better than the
random walk at short horizons for some countries; the results are however less favorable at
longer horizons.30
3.4 Commodity Prices and Other Predictors

Chen and Rogo¤ (2003) focus attention on commodity prices as a potential new macro-
economic fundamental for exchange rates. They focus on “commodity currencies”, that is
exchange rates for countries where primary commodities constitute a signi…cant share of
exports (i.e. Australia, Canada and New Zealand). Their main idea is that, typically, ex-
change rates are endogenously determined in equilibrium together with other macroeconomic
variables, so it is di¢ cult to predict exchange rate changes based on reduced-form models.
However, if it were possible to identify an exogenous shock to exchange rates, that would
cleanly predict exchange rate ‡uctuations. Chen and Rogo¤ (2003) argue that commodity
price changes act as “essentially exogenous”shocks for small open economies; the economies
29
Della Corte, Sarno and Sestieri (2010) use bilateral external imbalance measures, as opposed to the
global measure used in Gourinchas and Rey (2007), which summarizes the net foreign assets position of a
country vis-à-vis the rest of the world. They construct the measure of bilateral NXA based on Lane and
Milesi Ferretti’s (2007) database.
30
Alquist and Chinn (2008) use a proxy for NFA position based on the end-of-year US foreign asset and
liability data from the Bureau of Economic Analysis (BEA) and interpolated using US quarterly …nancial
account data from IFS.
16

with a large share of exports in primary commodities will typically experience exchange rate
appreciations when the price of their commodity exports increases. Measures of commodity
prices used in the literature to forecast exchange rate include commodity price indices (as in
Chen and Rogo¤, 2003, Chen, Rogo¤ and Rossi, 2010, and Ferraro, Rogo¤ and Rossi, 2011)
and oil prices (as in Bacchetta, van Wincoop and Beutler, 2010, and Rossi and Sekhposyan,
2011). Chen and Rogo¤ (2003) …nd in-sample empirical evidence in favor of commodity
prices as predictors of exchange rates; Chen, Rogo¤ and Rossi (2010) …nd that commodity
prices are not signi…cant out-of-sample predictors of exchange rates in quarterly data, and
Ferraro, Rogo¤ and Rossi (2011) …nd that they are in daily data.31
3.5 What Have We Learned?

The literature has considered a wide variety of predictors. Table 1 summarizes the main
predictors that have been used for out-of-sample exchange rate forecasting, and Table 2,
Panel A, provides an overview of which predictors have been used out-of-sample in several
of the papers we discussed. Overall, the empirical evidence is not favorable to traditional
economic predictors, except possibly for the monetary model at very long horizons and the
UIRP at short horizons, although there is disagreement in the literature. Both Taylor-
rule fundamentals and net foreign asset positions have promising out-of-sample forecasting
ability for exchange rates, although some papers question the robustness of the results. The
consensus in the literature is that these fundamentals have more out-of-sample predictive
content than traditional fundamentals; the disagreement in the literature is in the degree to
which these new fundamentals can explain the Meese and Rogo¤ puzzle.
4 Which Models To Choose?

Several models have been used in the literature in the attempt to forecast exchange rates.
Models can be divided into three representative groups: single-equation, multiple equations
and panel models. The models in each of these groups can be either linear or non-linear,
and may or may not allow for cointegration (i.e. by adding an error correction term) or time
31
Additional predictors that have been considered mainly for in-sample …t of exchange rate models include
order ‡ows (de…ned as the di¤erence between the number of buyer initiated transactions and the number of
seller initiated transactions). See Rime, Sarno and Sojli (2010) and Chinn and Moore (2012). As discussed
in Andersen et al. (2003, p. 59), order ‡ows have the potential drawback that we remain ignorant about
the macroeconomic determinants of high-frequency order ‡ows.
17

variation in the parameters. This section provides a guide to the various model speci…cations
that have been used in the literature, and evaluates both how the choice of the model may
a¤ect predictive ability as well as explain some of the di¤erences in …ndings across papers.
Table 2, Panel A, provides a convenient overview.
4.1 Single-equation, Linear Models

Meese and Rogo¤ (1983a, 1988) focus on models where exchange rate ‡uctuations are ex-
plained by the simple single-equation model:
0
Et (st+h st ) = 0 + 1 ft+h ; (9)
where the future, realized values of the fundamental ft+h are used. We refer to (9) as the
"single-equation, contemporaneous, realized fundamental model". The actual, rather than the
forecasted, value of the fundamentals is used as a predictor by Meese and Rogo¤ (1983a,b,
1988) to make sure that the lack of predictability of exchange rates is not due to poor
forecasts of the fundamentals. The parameters are estimated either by simple OLS or by
GMM (to deal with the endogeneity of the predictors). Meese and Rogo¤ (1983b) calibrate
the parameter in a grid to explore the robustness of their results to possible inconsistencies
in the parameter estimates. Cheung, Chinn and Pascual (2005) forecast exchange rates
using eq. (9) either by using calibrated parameter values (based on economic theory)32 or by
estimating eq. (9) via OLS (ignoring endogeneity issues). Similarly, Bacchetta, van Wincoop
and Beutler (2010) and Ferraro et al. (2011), among others, estimate eq. (9).33
To evaluate actual, ex-ante predictability of the fundamental, one might consider the
following models. The "single-equation, contemporaneous, forecasted fundamental" model
is:
0
Et (st+h st ) = 0 + 1 Et ft+h (10)
where Et ft+h is estimated based on information available up to time t, and the endogeneity
of the fundamentals requires instrumental variable estimation. Eq. (10) has been considered
by Meese and Rogo¤ (1983a) and Chinn and Meese (1995). In the "single-equation, lagged
fundamental model":
0
Et (st+h st ) = 0 + 1 ft ; (11)
32
For example, for PPP, they forecast future st+h st simply by using pt+h pt+h .
33
Ferraro et al. (2011) argue that the fundamental that they use (the rate of growth of commodity prices)
can be considered essentially exogenous; thus, they can consistently estimate the parameters by OLS.
18

contemporaneous fundamentals are directly used for forecasting the h-step-ahead rate of
growth of exchange rates.34 Note that OLS can be used to estimate the parameters, since
the regressors are lagged. Eq. (11) has been used by Wright (2008), Molodtsova and Papell
(2009, 2012), Molodtsova et al. (2010) and Rossi (2005a), among others. A slight modi…-
0
cation of eq. (11) is Et (st+h st ) =h = 0 + 1 ft , used by Gourinchas and Rey (2007) and
Della Corte et al. (2010).
The empirical evidence based on the single-equation linear model is mixed. Typically,
contemporaneous, realized fundamentals are not successful.35 Authors disagree on whether
using forecasted fundamentals improves the empirical evidence in favor of the monetary
model (Meese and Rogo¤, 1983a, …nd it doesn’t and Faust, Rogers and Wright (2003) …nd
that it does). The out-of-sample performance of lagged fundamentals depends on the pre-
dictor: it is poor for the monetary model with or without unemployment (see Rossi, 2006,
and Rossi and Sekhposyan, 2011), but successful for the Taylor-rule model (Molodtsova and
Papell, 2009, 2012, Molodtsova et al., 2010, and Giacomini and Rossi, 2010) and the net
foreign asset model (Gourinchas and Rey, 2007, and Della Corte et al., 2010). Ferraro et al.
(2011) …nd predictive ability using realized, contemporaneous commodity prices, though the
predictive ability weakens considerably using lagged commodity prices. Overall, we conclude
that, with the latter exception, it is the predictor that matters the most in determining the
strength of the predictability rather than the exact speci…cation of the single-equation linear
model.
4.2 Single-equation, ECM models

Since the work by Mark (1995), a model that has been widely used includes an error correc-
tion term. The Error Correction Model (ECM) assumes that there is a long run relationship
between the level of the exchange rate, st , and the level of the fundamentals, ft ; thus, in fore-
casting the rate of growth of the exchange rate, it includes a correction term that captures
34
Note that the forecast is based on a parameter directly estimated from a regression of the current rate
of growth of the exchange rate on the lagged value of the fundamental. See Marcellino, Stock and Watson
(2006) for a review of direct and iterated forecast methods.
35
See Meese and Rogo¤ (1983a,b) for the monetary model; Cheung, Chinn and Pascual (2005) for the
monetary, productivity and net foreign assets measures; and Bacchetta, van Wincoop and Beutler (2010)
for the monetary model augmented by unemployment and oil prices.
19

the long-run disequilibrium between the exchange rate and the fundamental levels:
P
h
0
Et (st+h st ) st+j = 0 + 1 (st ft ) (12)
j=1
To help intuition, consider the monetary model, eq. (3). Note that by substituting the
UIRP and PPP conditions in the relative money demand equation, we have: mt mt st
= (Et st+h st ) + (yt yt ), which leads to the popular approach of Mark (1995), where
Et st+h st = + [(mt mt ) (yt yt ) st ] : As long as < 1; exchange rates revert
back to their fundamental value ft (mt mt ) (yt yt ) over time. Thus, predictive
ability should be stronger at longer horizons. The cointegration vector parameter ( ) can
be calibrated – as in Mark (1995),36 Chinn and Meese (1995), Abhyankar et al. (2005),
Berkowitz and Giorgianni (2001) and Kilian (1999) –or estimated (typically by Stock and
Watson’s (1993) DOLS), either over the full-sample or recursively –as in Alquist and Chinn
(2008), Chinn and Moore (2012) and Cheung, Chinn and Pascual (2005).37
Positive evidence in favor of the ECM model at long horizons has been found by Mark
(1995), whereas Cheung, Chinn and Pascual (2005) and Alquist and Chinn (2008) …nd no
predictive ability. Note that the former calibrates the cointegration parameters whereas
the latter estimate them. On the other hand, using exactly the same ECM speci…cation,
Kilian (1999) and Groen (1999) …nd no predictive ability for the monetary model at long
horizons, whereas Rossi (2005a) does. We will investigate in detail the reasons behind the
disagreement over the predictive ability of the monetary model at long horizons in Sections
5.1, 6.4 and 6.5.38
4.3 Non-linear Models

Most of the literature focuses on linear models. Only a few papers focus on non-linear models,
including: non-parametric methods (locally-weighted regressions), as in Diebold and Nason
(1990), Meese and Rose (1991), Chinn (1991), Mizrach (1992) and Chinn and Meese (1995);
neural networks, as in by Qi and Wu (2003); or (exponential) transition autoregressive models
(ESTAR), as in Rapach and Wohar (2006). Again, the literature di¤ers regarding whether
actual, realized fundamentals or forecasted fundamentals are used –see Table 2.
36
Mark (1995) calibrates [1; ] = [1; 1] :
37
The parameters 0 ; 1 are estimated by OLS.
38
Amano and van Norden (1993) also use ECMs to predict the Canadian-US dollar real exchange rate
using world commodity price indices and interest rates; and Amano and van Norden (1995, 1998a,b) study
their cointegration properties.
20

The empirical evidence is not favorable to non-linear models. The vast majority of the
literature …nds that non-linear models forecast poorly,39 with some exceptions.40 This …nding
is consistent with the more general …nding that non-linear models …t well in-sample,41 but
fail in out-of-sample forecasting exercises (Terasvirta, 2006).
4.4 Time-Varying Parameter (TVP) Models

A special form of non-linearity may be induced by time-variation in the parameters. There
are several ways to deal with time variation in the parameters: either estimate a parametric
model where the parameters change over time according to a rule (as in the Kalman …lter
approach by Wol¤, 1987, and Schinasi and Swami, 1987; the Bayesian TVP model in Canova,
1993; Stock and Watson’s (1998) random walk coe¢ cient model; or the Markov Switching
model of Engel and Hamilton (1990) and Engel (1994)) or take averages across models, either
via a Bayesian Model Averaging (BMA) approach (Wright, 2008) or forecast combinations
(Timmermann, 2006).42
The "single-equation time-varying parameter" model used in Schinasi and Swami (1987)
and Wol¤ (1987) is:
0
st = t ft + ut ; (13)
t =G t 1 + K + Avt ;
where the parameter t changes over time according to an autoregressive process, K and
A are constants, and ut and vt are unforecastable shocks.43 The "random walk coe¢ cient
time-varying parameter" model (Stock and Watson, 1998) imposes G = I and K = 0:
39
Including Chinn (1991) and Chinn and Meese (1995) for the monetary model; Diebold and Nason (1990)
for univariate models; Mizrach (1992) for locally weighted regression model across several currencies; Qi and
Wu (2003) and Rapach and Wohar (2006).
40
Meese and Rose (1991) …nd signi…cant in-sample and out-of-sample predictability in the non-linear
monetary model. Satchell and Timmermann (1995) show that, although the squared forecast errors of
nonlinear models are higher than those of the random walk, non-linear models correctly predict a large
proportion of the sign of exchange rate changes for several countries.
41
Taylor and Peel (2000) and Kilian and Taylor (2003) are examples of in-sample estimation of non-linear
models and out-of-sample forecasting, respectively, in the case of real exchange rate models.
42
The di¤erence between BMA and forecast combinations is that BMA estimates the weights in the forecast
combination using Bayesian methods, whereas forecast combinations typically use equal weights or weights
estimated with frequentist methods.
43
Wol¤ (1987) estimated the TVP model using the Kalman …lter.
21

In the "Markov Switching" model, the relationship between exchange rates and fundamen-
tals depends on an unobservable variable. An example of the Markov Switching model is:
[st ; ft0 ]0 Rt N( Rt ; Rt ), where Rt = fR1 ; R2 ; :::; Rk g is an unobservable regime (or
state). For example, for k = 2, R1 could be a recession and R2 could be an expansion; in
this case, the relationship between exchange rates and fundamentals depends on the state
of the business cycle. The regime evolves stochastically, according to a law of motion (see
Engel and Hamilton, 1990, for details on the estimation and the model).
The empirical evidence is mixed. Some papers …nd out-of-sample forecast improvements
over the random walk –see Schinasi and Swamy (1987) and Wol¤ (1987) for TVP models,44
and Engel and Hamilton (1990) and Engel (1994) for a Markov Switching univariate model.
In other cases, the forecast improvements are unclear: Rossi (2006) …nds that a random walk
coe¢ cient TVP model outperforms the random walk for only one country; Bacchetta, van
Wincoop and Beutler (2010) instead conclude that time variation cannot explain the Meese
and Rogo¤ puzzle.45
4.5 Multivariate Models

Many di¤erent types of multivariate models have been used in the literature, typically gen-
eralizations of the single-equation models discussed above. Let Yt [st ; ft0 ]0 : Among the
multivariate models, the following have been prominent.
(a) VARs. VARs have been considered in Meese and Rogo¤ (1983a,b):
A (L) Yt = ut ; (14)
where A (L) = I A1 L ::: Ap Lp and L is the lag operator. A special version of VARs are
Bayesian VARs (BVARs), that is VARs estimated with a large number of variables imposing
some Bayesian shrinkage for the parameters, which otherwise would be very imprecisely
estimated in the small samples typically available to researchers. Meese and Rogo¤ (1983a,b)
…nd that VARs with monetary fundamentals do not improve over the random walk. The
BVARs with a large number of exchange rates in Carriero, Kapetanios and Marcellino (2009)
outperform the random walk even at short horizons. Note that the BVAR considered by
44
although the former did not test their signi…cance and the latter …nds signi…cant evidence for one country
out of three. Cheung and Erlandsson (2005) …nd in-sample empirical evidence in favor of a Markov Switching
model for exchange rates.
45
Although Chinn (2009) and Giannone (2009) debate their conclusion.
22

Carriero, Kapetanios and Marcellino (2009) does not include any economic fundamentals
and therefore does not shed light on which fundamentals economists should focus on.
(b) Factor models. The "factor ECM model" proposed by Engel, Mark and West
(2009) is:
Et (si;t+h si;t ) = i + h ( i0 't si;t ) ; (15)
where 't is the factor extracted from the explanatory variables Xt , (Xt = 't +ui;t , where the
explanatory variables are a panel of exchange rates), i denotes the country, and the number
of countries considered is large.46 Engel, Mark and West (2009) …nd that, in some cases,
the factor model improves forecasts at long horizons but does not improve short-horizon
forecasts.47
(c) VECMs. The single-equation ECM model, eq. (12), is a simpli…cation of the tradi-
tional "multi-equation VECM model":
0
st+1 = 0 + 1 (st ft ) + 2 (L) st 1 + 3 (L) ft 1 (16)
where the short run dynamics is eliminated, and 1 L. The empirical evidence on
VECMs is mixed: some papers …nd positive evidence (i.e. MacDonald and Taylor, 1993,
for the monetary model; Clarida and Taylor, 1997, for forward rates) while others are more
negative (Rapach and Wohar, 2002, and Diebold, Gardeazabal and Yilmaz, 1994).48 Overall,
the literature suggests that single-equation ECM speci…cations are preferable to VECMs
because the short-run dynamics of exchange rates is di¢ cult to estimate (cfr. Cheung,
Chinn and Pascual, 2005, p.1156)
(d) Multivariate time-varying parameter models. Multivariate models can also
have time-varying parameters: a multivariate version of the time-varying parameter model
(13) has been used in Canova (1993).49 Due to the complexity, these models are estimated
by Bayesian methods. The empirical evidence shows that multivariate TVP models may
provide forecast improvements over the random walk (see Canova, 1993, for the multivariate
46
Eq. (15) may include additional control variables, such the deviation of Taylor rules, monetary or PPP
fundamentals from the current exchange rate of the country.
47
They also consider the alternative speci…cation: Et si;t+h = i0 Et 't+h .
48
Diebold, Gardeazabal and Yilmaz (1994) reconcile their negative …nding with the fact that powerful
cointegration tests reject cointegration. See also Baillie and Bollerslev (1989, 1994) for in-sample tests of
cointegration among exchange rates.
49
More in detail, Canova’s (1993) model is: Bt (L) Yt = ut , where ut j=t N (0; V ), =t is the information
set, and B t = GB t 1 + K + Avt ; where B t is the vector containing all the parameters in Bt (L) and
vt (0; t ).
23

TVP model with interest rates, and Clarida, Sarno, Taylor and Valente, 2003, for the Markov
Switching VECM model with forward rates).
(e) Bayesian Model Averaging (BMA). An alternative way to exploit predictability
from many regressors is via Bayesian model averaging, which combines forecasts from com-
peting models using weights that are estimated by posterior probabilities. Wright (2008)
considers BMA with predictors that include both traditional predictors as well as stock
prices, dividend yields and the current account. He …nds that, for most currencies, BMA
with su¢ ciently high shrinkage produces forecasts that are better than the random walk,
although the magnitude of the improvements is marginal.
4.6 Panel Models

Several panel models have been estimated in the literature. The "panel ECM" considered
by Mark and Sul (2001), Groen (2005), Engel, Mark and West (2007), Cerra and Saxena
(2010) and Rapach and Wohar (2004) is:
Et (si;t+h si;t ) = h (fi;t si;t ) ;
where the error term contains both an individual, time-invariant component, an aggregate,
time variant component, and an individual, time variant component.50 Mark and Sul (2001),
Groen (2005), Engel, Mark and West (2007) and Cerra and Saxena (2010) impose a known
cointegrating parameter .51 Rapach and Wohar (2004) estimate the cointegrating parame-
ters recursively, but unfortunately do not compare the models’ forecasts with the random
walk. The "panel, contemporaneous realized fundamental model" used by Cerra and Saxena
(2010) is: Et si;t+h = fi;t+h , where fi;t+h is the actual, realized value of the fundamen-
tal. The "panel, lagged fundamental model" used by Adrian, Etula and Shin (2011) is:
Et si;t+h = fi;t , where fi;t is the lagged value of the fundamental.
The empirical evidence suggests that panel ECMs are quite successful for the monetary
model (see Mark and Sul, 2001; Groen, 2005; Engel, Mark and West, 2007; and Cerra and
Saxena, 2010) and for funding liquidity fundamentals (Adrian, Etula and Shin, 2011). They
are less successful for PPP fundamentals (Mark and Sul, 2001).
50 1
Pt
Its forecasts are: h (fi;t si;t ) + i + t j=1 j , where i is the individual, time-invariant component
and t is the aggregate, time-variant component.
51
In particular, Groen (2005) imposes knowledge on the number of cointegrating vectors as well as the
parameters of the common cointegrating vector, and only re-estimates the intercept and the coe¢ cients of
short-run dynamics.
24

Among the model speci…cations considered in the literature, the least successful are non-
linear speci…cations, and the most successful are linear speci…cations. Among the linear
models, the single-equation ECM and the panel ECM models are the most successful at long
horizons, although there is disagreement among researchers about the degree of robustness
of the results. Typically, but not always, for single-equation linear models the predictor
choice matters more than whether the researcher uses contemporaneous, realized or lagged
fundamentals; the linear monetary model does not perform well at any horizons, whereas
Taylor-rule fundamentals and net foreign assets are successful predictors at short-horizons.
Models with time-varying parameters show some degree of success, although most of the
favorable empirical evidence is based on studies in the late 1980s-early 1990s, and the most
recent analyses report less success (although with slightly di¤erent speci…cations). Among
the multivariate models, the most successful speci…cation is the panel ECM, although there
is some evidence in favor of BVARs, BMA and factor ECM models.
The next section will investigate why, among researchers using the same model, there is
substantial disagreement across empirical …ndings by examining in details other important
dimensions in which the studies di¤er.
5 Which Data To Use?

Existing studies in the literature di¤er considerably depending on the characteristics of their
data. In particular, they di¤er depending on the data transformation they perform, the
countries they study, the sample they use, the frequency of the data, and whether the data
are fully revised or real-time. This section reviews what choices have been made in the
literature and what we know about how they potentially a¤ect predictability; in particular,
we include a discussion that revisits the empirical …ndings in the previous sections in this
light. Panel B in Table 2 overviews data choices in selected papers in the literature.
5.1 End-of-sample versus Filtered Data and Calibrated Parame-

ters
When conducting out-of-sample forecast evaluation, it is important to make sure that the
information contained in future data is not trivially used to estimate the models’parameters.
25

To clarify, consider a forecaster estimating the model at time t and producing the h-step-
ahead forecast for time t + h. It is important that when the forecaster produces the forecast
at time t, the information contained between times t and t + h (or later) is not used directly
or indirectly in the estimation, otherwise the model has an unfair advantage relative to the
random walk. In other words, the model would be given the advantage of “peeking” at
data after time t, which would not have been available to the forecaster in real time. One
example when this might happen is when the data are de-trended over the full sample or
de-trended by any two-sided …ltering procedure. For example, when using the output gap to
forecast exchange rates, it is important that the output gap, usually de…ned as the di¤erence
between output and a linear or a quadratic trend, be de-trended using only information
available at time t: Seasonal adjustment is another example where the treatment of the data
is important. All the raw data used in Meese and Rogo¤’s (1983a) study were seasonally
unadjusted: they performed the seasonal adjustment procedure themselves on the raw data.
The reasons are twofold: using raw data allows researchers to seasonally adjust the data in a
consistent way across di¤erent series; furthermore, it allows them to avoid using information
that was not available at the time the forecast was made (for example, it is preferable to
use a one-sided moving average seasonal adjustment …lter, as opposed to a two-sided one,
as forecasts based on two-sided …ltered data implicitly use information which would have
not been available at the time of the forecast). As a third example, one should focus on
forecasting actual exchange rates, and not Hodrick-Prescott …ltered exchange rates. As a
fourth example, using calibrated fundamentals (e.g. linear combinations of macroeconomic
predictors, where the coe¢ cients are calibrated rather than estimated) is another example
of "ad-hoc" …ltering, which could potentially give the model an unfair advantage since it is
unclear how the parameters are calibrated (in particular, what data inspired the calibrated
parameters). As a …fth example, sometimes, auxiliary parameters are imposed in an "ad-
hoc" manner, and it is unclear whether results are robust to such choices. Consider the case
of a researcher using BMA techniques and shrinking the parameter towards the random walk
(that is, imposing a rule to make the parameter smaller); if the rule is ad-hoc, it is unclear
whether peeking at the whole dataset might have in‡uenced the degree of shrinkage. Other
examples (including the choice of the estimation sample or the use of real-time data) are
discussed in detail below.
Let us critically overview how these "ad-hoc" procedures may a¤ect the favorable …nd-
ings of predictive ability reviewed in the previous sections. Typically, in the ECM estimation
26

literature, researchers impose a known cointegrating parameter.52 The predictability found
in VECMs with monetary fundamentals also typically relies on the cointegrating parame-
ter estimated over the full sample:53 Cheung, Chinn and Pascual (2005) and Alquist and
Chinn (2008) show that the empirical evidence in favor of long horizon predictability of the
monetary model is much weaker or completely disappears after estimating the cointegrating
parameters. Thus, some of the disagreements over the empirical performance of the mon-
etary model at long horizons can be explained by di¤erences in models’ speci…cation: the
model performs better by imposing an ad-hoc calibrated cointegration parameter than by
estimating it.54 In some panel analyses (i.e. Adrian, Etula and Shin, 2011) the fundamentals
are de-trended over the full sample prior to forecasting as well. The favorable evidence for
the BVAR relies on choosing the degree of Bayesian shrinkage, and it is unclear how di¤erent
degrees of shrinkage may a¤ect the results.
5.2 Forecasted versus Realized (or Ex-post) Fundamentals

The previous section highlights that the same model can be estimated either using forecasted
or realized fundamentals. The latter also give an unfair advantage to the fundamentals-
based economic model. However, it is important to note that the original Meese and Rogo¤
(1983a,b) analysis used future realized fundamentals, as opposed to fundamentals known
at the time of the forecast, to predict exchange rates because they were proving a negative
result. The result that Meese and Rogo¤ (1983a,b) wanted to prove is that no fundamental
forecasts exchange rates better than a random walk; by showing that not even realized,
h-period-ahead fundamentals were capable of predicting exchange rates, they made their
…ndings really stark. In other words, if forecasted h-period-ahead fundamentals had no
predictive content for h-period-ahead exchange rates, that might be due either to the lack of
predictive content or to the fact that forecasts of fundamentals are poor. By demonstrating
that even when using realized fundamentals it is not possible to improve exchange rate
forecasts, Meese and Rogo¤ (1983a,b) were able to prove that the lack of predictive content in
52
For example, Mark (1995) and Chinn and Meese (1995). Typically, papers in this literature also impose
knowledge on the number of cointegrating vectors.
53
For example, MacDonald and Taylor (1994) …nd that the sticky-price monetary model estimated via
cointegration outperforms the random walk out-of-sample; however the cointegrating vector is estimated
over the entire sample, generating forecasts that incorporate future realized values and hence they are not
truly ex-ante.
54
In fact, Groen (1999) argues that the cointegration relationship between exchange rates and monetary
fundamentals is unstable and disappeared in the later sample.
27

the fundamentals is really the cause of the problem. Table 2 provides an overview on the use
of forecasted versus realized fundamentals in the literature. As noted in the previous section,
typically for single-equation linear models, the predictor choice matters more than whether
the researcher uses contemporaneous, realized or lagged fundamentals. For panel models,
Cerra and Saxena (2010) demonstrate that the use of realized versus lagged fundamentals
does not matter either.
5.3 Countries and Samples

Existing studies in the literature di¤er with respect to the countries they consider as well as
the sample. Most studies focus on bilateral exchange rates versus the US dollar, although
there are exceptions.55 Typically, the sample starts in 1973 to avoid the period of …xed
exchange rates and typically the end of the sample is constrained by data availability; for
example, individual European countries’exchange rates are available only prior to the Euro
uni…cation. More recent studies have replaced the exchange rate of the countries that are now
part of the European Union with the euro-dollar exchange rate; its value can be back-casted
to the period before the uni…cation as well, although the e¤ects of back-casting are unclear.
A clear message in the literature is that models and predictors that forecast exchange rates
well in one country do not necessarily provide competitive forecasts for other countries (see
Cheung, Chinn and Pascual, 2005).
5.4 Frequency
Existing studies in the literature also di¤er with respect to the frequency of the data they
consider, ranging from low frequency, yearly data, to quarterly, monthly, and even weekly,
daily or very high-frequency data (such as those available at 30-minute intervals). There
is clearly a trade-o¤ between the frequency of the data and the span of the data, as lower
frequency data (which are perhaps more informative regarding long-run trends in the data)
are by construction available only for shorter samples.56 Typically, studies interested in the
long-term forecastability of exchange rates focus on quarterly data (i.e. Mark, 1995). There
are no striking discrepancies between empirical results of studies that focus on monthly data
55
For example, Meese and Rogo¤ (1983a,b) consider also the trade-weighted exchange rate.
56
E.g. longer spans of data have been considered by Rapach and Wohar (2002) at the price of restricting
attention to yearly frequencies and a smaller sample. See also Taylor (2002) on using longer span of data
for the analyzing real exchange rates.
28

relative to those that focus on quarterly data; some of the studies that focus on daily data (for
example, studies that use macroeconomic news announcements as predictors)57 do generally
…nd positive evidence for the model, but typically they are estimated in-sample, so the real
di¤erence seems to be whether the models are compared in-sample versus out-of-sample.
We conclude that the frequency of the data does not typically a¤ect …ndings of predictive
ability.58
5.5 Data Revisions

Typically, researchers evaluate the out-of-sample forecasting ability of their models using
data available at the time of their study. However, the data that a researcher has available
in 2000Q1 for, say, GDP in 1998Q1 is not the same as that available to a researcher in
1998Q1. The reason is that data are continuously revised by statistical agencies. The
revisions include those due to statistical agencies acquiring additional source information to
update their initial estimates, changes in the aggregation methods (…xed to chain weighting),
changes in base years used for calculating real variables, or changes in de…nitions of the
concept being measured. Financial data, such as interest rates and exchange rates, are never
revised, but GDP, money aggregates and in‡ation are. A real-time dataset collects vintages
of data that were actually available to researchers at each point in time, before data revisions
were applied to the data to obtain the …nally revised data available at the end of the sample
(or current vintage data). The conventional wisdom is that, for example, the use of revised
data may lead researchers to include regressors that had little predictive content in real-time
and may exaggerate the forecasting ability relative to predictors that were actually available
at the time (see Koenig, Dolmas and Piger, 2003, and Croushore and Stark, 2001,2003).59
57
Macroeconomic news announcements are de…ned as the di¤erence between expected and realized (“an-
nounced”) macroeconomic fundamental values. Expectations are typically measured by surveys. For exam-
ple, Andersen, Bollerslev, Diebold and Vega (2003), Ehrmann and Fratzscher (2005), Faust, Rogers, Wang
and Wright (2007) and Fratzscher (2009) study whether, in-sample, exchange rates react in a very short
window around the announcement.
58
One exception might seem Ferraro et al. (2011). However, even in their case, the main reason for the
predictive ability is not just the use of daily data, but the use of both daily data and commodity prices as
fundamentals.
59
Note that the dependent variable (the exchange rate) is a …nancial variable and is never revised. Thus,
unlike forecasting output growth, where the researcher needs to determine whether to forecast the …rst
revision, the second revision or the latest revised data, which value to forecast is an irrelevant issue when
forecasting exchange rates.
29

The empirical evidence shows that the …ndings in favor of the long-horizon predictive
ability of the monetary model are not robust to using real-time data. In particular, Faust,
Rogers and Wright (2003) show that the favorable evidence of long-horizon predictive ability
is present only over a two-year window of data vintages around those used in Mark (1995):
most of the predictive ability eventually disappears after data revisions. Interestingly, they
also …nd that the economic model performs better using real-time data than …nally revised
data. On the other hand, the empirical evidence shows that the predictive ability of Taylor-
rule models is robust to using real-time data –see Molodtsova et al. (2010).

Data transformations (such as de-trending, …ltering and seasonally adjusted data) may cru-
cially a¤ect predictive ability, and may explain why, for the same models, some researchers
do …nd predictive ability while others do not. Cheung, Chinn and Pascual (2005) and Alquist
and Chinn (2008) …nd that the empirical evidence in favor of the single-equation ECM mon-
etary model is much weaker or completely disappears after estimating the cointegrating
parameters. The evidence in favor of the panel ECM model also relies on calibrated cointe-
gration parameters, and it is unclear whether the …ndings would be robust to estimating the
parameters; in addition, some panel ECM studies rely on data that have been previously de-
meaned over the full sample, which may create an unfair advantage for the economic model
over the random walk. Another important factor that, for some fundamentals, may weaken
predictive ability is using real-time, rather than realized, data; this is the case for monetary
fundamentals but less of a concern for Taylor-rule fundamentals. For a given model and
predictor, predictive ability also depends on the choice of the country. On the other hand,
the frequency of the data and whether the realized or the forecasted fundamental is used do
not seem to a¤ect predictability: the monetary model’s forecasts do not beat a random walk
either way. However, there are exceptions.60
60
Sometimes forecasts perform better using real-time forecasts of future fundamentals instead of actual
future fundamentals, as in Faust, Rogers and Wright (2003). Ferraro et al. (2011) …nd that predictability is
stronger using realized (as opposed to lagged) fundamentals.
30

6 Which Forecast Evaluation Methods To Choose?
This section provides an overview of the methods that have been used in the literature to
assess the predictive content of macroeconomic fundamentals for exchange rate forecasts as
well as important issues to keep in mind when evaluating the predictive content. We also
critically revisit the empirical …ndings in the previous sections in the light of these issues.
Table 2, Panel C, overviews the forecast evaluation methods used in selected papers.
6.1 Choice of Benchmark Model

The majority of studies compare the out-of-sample forecasting performance of the predictors
with those of a random walk without drift ("RW"), as the latter has been shown to be the
best predictor of exchange rates since Meese and Rogo¤ (1983a,b) – see Table 2, Panel C.
According to the random walk without drift, the best predictor of exchange rates tomorrow
is the exchange rate today.61 Thus, exchange rate changes are completely unpredictable:
Et st+h st = 0
An alternative benchmark is the random walk with drift ("RWWD"), according to which
exchange rate changes are predictable but independent of other macroeconomic variables:
Et st+h st = 6= 0.
It is important to note that the e¢ cient market hypothesis does not imply that exchange
rate changes should be unpredictable. That is, the Meese and Rogo¤ (1983a,b) …nding that
the random walk provides the best prediction of exchange rates should not be interpreted
as a validation of the e¢ cient market hypothesis. The e¢ cient market hypothesis means
that bilateral exchange rate is the market’s best guess of the relative, fundamental value of
two currencies based on all available information at that time. The e¢ cient market hypoth-
esis does not mean that exchange rates (like any asset prices) are unrelated to economic
fundamentals, nor that exchange rates should ‡uctuate randomly around their past values.
Consistently, across papers, the random walk is used as the benchmark. Typically, when
the random walk with drift forecasts are reported, they are worse than those of the random
walk without drift. This may explain some puzzling results in the literature. For example,
among papers that study non-linear models, Meese and Rose (1991) found more empirical
evidence in favor of the nonlinear model, whereas Chinn (1991) and Chinn and Meese (1995)
61
When not speci…ed otherwise, random walk means the random walk without drift.
31

found the least. However, the benchmark in the former paper is the random walk with drift,
whereas the benchmark in the latter two papers is the random walk without drift.
6.2 Choice of Forecast Horizon

Forecast evaluation also requires the choice of a forecast horizon, h. Forecast horizons typi-
cally considered in the literature range from short horizons (typically 1 month or 1 quarter
ahead, depending on whether data are monthly or quarterly) to long horizons (horizons up
to 4-5 years: for example, Mark (1995) considers forecast horizons up to 16 quarters, and
Cheung, Chinn and Pascual (2005) up to 20 quarters).
The choice of the forecast horizon is important since it is possible that the models’pre-
dictive ability may depend on it. Typically, the empirical evidence in favor of fundamentals’
predictability appears at horizons that di¤er depending on the predictor. For example, for
monetary fundamentals, most studies agree on lack of short-horizon predictability (i.e. Che-
ung, Chinn and Pascual, 2005); however, there is evidence of long horizon predictability at
the 3-4 year horizon (Mark, 1995). The predictability of Taylor-rule fundamentals is gener-
ally found at short horizons (Molodtsova and Papell, 2009), whereas the predictability of net
foreign assets models is found at both short and long horizons by some studies (Gourinchas
and Rey, 2007), and only at short horizons in some other studies (Alquist and Chinn, 2008).
6.3 Choice of Forecast Methodology

Assessing the forecasting ability of a model requires several choices. First, which forecasting
method to use. The literature has been focusing mainly on rolling or recursive window
forecasting schemes (see West, 1996), where parameters are re-estimated over time using a
window of recent data –see Section 2. Shorter estimation window sizes allow the parameter to
adapt more quickly to structural changes; on the other hand, the parameter is less e¢ ciently
estimated in a smaller sample.62 The choice of how to split the sample between in-sample
and out-of-sample periods as well as the choice of the rolling window size varies across papers.
For example, among papers using a rolling window scheme, Meese and Rogo¤ (1983a) use
a window of 93 observations, Chinn (1991) uses 48, Qi and Wu (2003) use 216, Cheung et
al. (2005) consider 42 and 59, van Dijk and Franses (2003) use 128, Clark and West (2006,
2007) and Molodtsova and Papell (2009) use 120 observations. See Table 2, Panel C, for
additional details.
62
See e.g. Rossi (2011).
32

Rogo¤ and Stavrakeva (2008) and Inoue and Rossi (2012) examine the robustness of
Taylor-rule and panel ECM monetary models and …nd that their performance changes sub-
stantially depending on the estimation window size. Thus, it is not the case that the pre-
dictive ability is present no matter which window size is used, as pointed out in Rogo¤ and
Stavrakeva (2008); on the other hand, there are window sizes for which the predictive ability
is signi…cant, as discussed in Inoue and Rossi (2012). Regarding the net foreign asset model,
Rogo¤ and Stavrakeva (2008) …nd that its performance is less sensitive, although still not
entirely robust, to the choice of the window size.
6.4 Choice of Evaluation Methods

The forecast evaluation process requires two main choices: which loss function to use to
evaluate the forecast, and which test statistic to use to assess signi…cance. Regarding the
choice of loss function, researchers typically evaluate models’out-of-sample forecasting per-
formance according to their mean squared forecast error ("MSFE") or root mean square
forecast error ("RMSFE"), as in Meese and Rogo¤ (1983a,b, 1988). The former is simply
the average, over the out-of-sample period, of the squared forecast errors of a model; the
latter is its square root – see Section 2. Typically, researchers focus on either the RMSFE
di¤erence of competing forecasts or Theil’s (1996) U statistics, which is the ratio of model
RMSFEs to the RMSFEs of the benchmark model.63 Note that this is only one of the
available metrics/loss functions that can be used. Alternatively, researchers have used mean
absolute errors ("MAE", as in Meese and Rogo¤, 1983a) and asymmetric loss functions (Ito,
1990, West, Edison and Cho, 1993). Di¤erent loss functions give di¤erent weight to the devi-
ations of the forecast from the target; for example, the MSFE gives equal weight to over- and
under-predictions of the same magnitude.64 Alternatively, the forecast evaluation exercise
could target: (i) the direction of prediction, as in Engel (1994), Cheung et al. (2005) and
Cerra and Saxena (2010); (ii) a utility-based measure, as in West, Edison and Cho (1993),
Rime, Sarno and Sojli (2010), Abhyankar et al. (2005) and Della Corte et al. (2010); or (iii)
the whole predictive density or interval forecasts, as in Diebold, Hahn and Tay (1999), Inoue
and Rossi (2008) and Wang and Wu (2009). The direction of change statistic calculates the
proportion of forecasts that correctly predict the direction of change of the exchange rate.65
63
While there are other forms of Theil U statistic, this is the one commonly used in the exchange rate
literature.
64
See Elliott and Timmermann (2008) for a review of asymmetric loss functions.
65
See Pesaran and Timmermann (1992) for a test for the direction of prediction.
33

As opposed to the RMSFE criterion, it focuses on whether the sign of the forecast change
is correct, and, unlike the RMSFE, is not in‡uenced by the distance between the forecast
and the actual realization. In fact, it is theoretically possible that a model could forecast
perfectly the direction of change in all periods and yet forecast worse than the random walk
according to the RMSFE criterion; this may happen, for example, if the model consistently
over-predicts the magnitude of the change. For example, Satchell and Timmermann (1995)
theoretically prove that, whereas in linear models the probability of predicting the sign of a
stochastic variable decreases when the MSFE increases, this monotonic relationship between
the MSFE and the probability of correctly predicting the sign of the target variable does not
need to hold in general non-linear models. Utility-based measures are typically interpreted
as the transaction fee that a professional money manager can charge for providing estimates
of the economic fundamental model to an investor that uses the benchmark model. The
evaluation of the correct speci…cation of density forecasts goes beyond point forecasts by
evaluating the whole predictive density.66
The signi…cance of superior forecast performance is typically assessed via out-of-sample
predictive ability tests or in-sample Granger-causality tests. Both provide important insights,
and are used for di¤erent goals. The former are used for assessing whether predictors would
have improved exchange rate predictions in forecasting environments that mimic as closely
as possible the one faced by forecasters in practice, and have been used in Meese and Rogo¤
(1983a,b). The latter are in-sample tests on whether the lagged predictor has signi…cant
explanatory power for exchange rates over the full sample, and have been used for example
in Andersen et al. (2003).67 It is important to keep in mind that the former is a much
more challenging exercise than the latter: predictors which pass the latter may still not have
predictive ability in a truly out-of-sample forecasting exercise. Indeed, the Meese and Rogo¤
66
It typically relies on the Probability Integral Transform (see Diebold, Hahn and Tay, 1999, and Corradi
and Swanson, 2006). See also Rossi and Sekhposyan (2012) for tests robust to instabilities.
67
A di¤erent evaluation method is used in Engel and West (2005). They focus on evaluating whether
exchange rates Granger-cause fundamentals in-sample (rather than vice-versa). The reason is because they
show that, if fundamentals have a unit root and the discount factor is near one, the net present value
condition implies that exchange rates are near–random walks. Since the present value condition implies that
exchange rates should predict fundamentals, they tested whether exchange rates Granger-cause fundamentals
in-sample. While this is an implication of their framework, and therefore a clever way to provide indirect
empirical evidence to substantiate their theory, it is nevertheless not a necessary condition as exchange
rates might Granger-cause future interest rates either because of the net present value condition or because
interest rates are set by Central Banks to mitigate exchange rate ‡uctuations. Engel and West (2004) study
the implications of a discount factor near unity for exchange rate volatility.
34

puzzle is the …nding that, although fundamentals are signi…cant predictors of exchange rates
in-sample, their out-of-sample predictive ability is not superior to that of the random walk. A
version of Granger-causality test robust to instability is Rossi (2005b); unlike the traditional
Granger-causality tests, it detects predictive ability even if it appears only in a sub-sample,
or if the predictive relationship changes over time. The latter test has been used, for example,
in Chen, Rogo¤ and Rossi (2010).
Traditional tests of out-of-sample predictive ability can be distinguished between absolute
tests and relative tests of forecast evaluation. The former evaluate properties that optimal
forecasts of a model should have, such as unbiasedness and uncorrelatedness.68 The latter
evaluate which, among competing models, forecasts the best, and include the tests proposed
by Diebold and Mariano (1995), West (1996), Clark and McCracken’s (2001) ENCNEW,
Giacomini and White (2006), Clark and West (2006, 2007), among others. The Meese and
Rogo¤ puzzle concerns tests of relative predictive ability, since Meese and Rogo¤ (1983a,b)
compare the forecasting ability of fundamental-based exchange rate models with the random
walk. While the tests for relative forecast performance developed in the literature are typi-
cally applied to MSFE di¤erences between models, there is an important di¤erence among
them: on the one hand, West (1996), Clark and McCracken (2001) and Clark and West
(2006,2007) test out-of-sample whether the benchmark model is equivalent to the compet-
ing model in population; instead, Diebold and Mariano (1995) and Giacomini and White
(2006) test whether two models’forecasting ability is the same. I.e. the former test, using
out-of-sample forecasts, whether, say, = 0 in Et (st+h st ) = ft ; the latter test whether
the forecasts of the fundamental model and that of the random walk are equivalent. Thus,
the latter approach might be useful when the researcher is interested in evaluating forecasts,
and the former when (s)he is interested in evaluating models in population. The di¤erence
between the two approaches is that, in nested models, the sample MSFE from the larger
model is expected to be greater than that of the small model even when, in population, the
two models have the same predictive ability, since the larger model introduces noise into
its forecasts by estimating parameters that are useless in forecasting –see Clark and West
(2006). Therefore, a …nding that the small model has a smaller MSFE does not necessarily
mean that the additional predictors present under the larger models are not useful for fore-
casting. The former tests take this into account in evaluating the relative predictive ability
of competing models by re-centering the MSFE di¤erential in the test statistic by a term
68
See Mincer and Zarnowitz (1969), West and McCracken (1998), and Patton and Timmermann (2006,
2007).
35

that accounts for parameter estimation error and hence test whether, after adjusting for
parameter estimation error, the larger model outperforms the benchmark – but parameter
estimation error is exactly one of the reasons why there is a di¤erence between in-sample …t
and out-of-sample forecasts. Since the tests rely on di¤erent assumptions and null hypothe-
ses, they have di¤erent power properties. See West (2006) and Diebold (2012) for insightful
discussions. The majority of the papers in the exchange rate literature focus on the Diebold
and Mariano (1995) and West (1996) tests (hereafter "DMW") or the Clark and West (2006,
2007) tests (hereafter "CW"). Table 2, Panel C, provides an overview of which tests have
been used in the literature.
The main conclusion that emerges from the literature is that the choice of the evaluation
method matters, and a lot! For example, the use of di¤erent test statistics may explain
the contradicting evidence on the empirical validity of UIRP. Typically, the majority of the
studies that …nd lack of predictability for interest rate di¤erentials either focus on RMSFEs
or on the Diebold and Mariano (1995) test. Clark and West (2006) …nd that, based on the
Diebold and Mariano (1995) and West (1996) tests, there is little evidence that UIRP beats
the random walk; however, UIRP produces better forecasts than the random walk according
to the Clark and West (2006) test. Similarly, using the Clark and West (2006) test, Alquist
and Chinn (2008) conclude that UIRP can signi…cantly outperform the random walk at long
horizons.
The choice of the forecast evaluation method may also shed some light on the contro-
versy over the empirical evidence on long horizon predictability of monetary fundamentals.
On the one hand, Kilian (1999) and Berkowitz and Giorgianni (2001) dispute the …nding
of predictive ability of the monetary model at long horizons due to issues with the boot-
strap implementation of the test and the fact that high persistence would distort inference.
In particular, Berkowitz and Giorgianni (2001) and Berben and van Dijk (1998) show, via
Monte Carlo simulations, that predictive ability measures of two completely unrelated time
series could display large R-square and DMW statistics at long horizons if they are both
highly persistent, a …nding reminiscent of spurious regression. Kilian (1999) instead argues
that, when bootstrapping the DMW statistic, it is important that the critical values of the
statistic be calculated under the null hypothesis of the absence of a drift in the random walk,
otherwise the bootstrap model is not consistent with the model used as the benchmark. On
the other hand, Rossi (2005a) shows that the high persistence in exchange rates and funda-
mentals biases the estimation of the economic models, but not the random walk: when this
bias outweighs the bene…ts from exploiting economic predictors, the random walk forecasts
36

better.69 Her …ndings show that the poor forecasting ability of economic models does not
imply that the models are not a good description of the data.70
The choice of the evaluation metric may also be important in explaining di¤erences in
empirical results across studies. While Meese and Rogo¤ (1983a,b) show that their negative
…nding is robust to using either mean squared forecast errors or mean absolute errors, several
papers note that other metrics are more successful in …nding out-of-sample evidence in favor
of macroeconomic predictors: Engel (1994) and Cheung, Chinn and Pascual (2005) …nd
that the direction of change statistics provides more empirical evidence that models can
outperform the random walk than MSFE comparisons; Abhyankar et al. (2005) …nds similar
results for utility-based measures.
6.5 Choice of the Forecast Sample

The sample utilized for out-of-sample forecast evaluation di¤ers substantially across studies,
not only because the studies have been performed at several points in time and di¤er in the
overall samples they consider, but also because they di¤er in the choice of the window size
used for estimation. Clearly, in the presence of instabilities, the models’performance could
potentially be sensitive to the sample used for forecast evaluation.
Again, the literature is very clear on the following: the choice of the forecast sample
matters enormously! In fact, one of the most interesting results in the literature concerning
the robustness of results to the forecast period is the lack of robustness. For example, long
horizon predictability of the monetary model is very sensitive to the forecast period: Kilian
(1999) and Groen (1999) …nd that the long horizon predictability of the monetary model
disappears once one extends Mark’s (1995) sample. Cheung, Chinn and Pascual (2005)
consider two out-of-sample periods and note that the superior performance of a particular
combination of monetary fundamental, model speci…cation and country combination does
not necessarily carry over from one out-of-sample period to the other. Rogo¤ and Stavrakeva
(2008) note that neither the out-of-sample forecasting ability of panel VECM models nor
that of Taylor rules are robust to the choice of the out-of-sample forecast period. Giacomini
and Rossi (2010) note substantial changes over time in the predictive ability of Taylor-rule
and UIRP fundamentals relative to the random walk.
The lack of robustness of results across sample periods is consistent with the possibility
69
This happens even if the economic model is the true data generating process.
70
A further limitation of traditional forecast evaluation tests is that they typically rely on stationarity
assumptions, and are not robust to the presence of instabilities –see Section 6.5.
37

that the parameters be unstable: in fact, Rossi (2006) shows that parameter instabilities
are empirically important in exchange rate prediction. This suggests that it is crucial to
consider evaluation methodologies that are robust to instabilities, such as Giacomini and
Rossi’s (2010) test, rather than traditional methods.71 For example, using their proposed
test, Giacomini and Rossi (2010) uncover a sharp worsening in the forecasting ability of UIRP
in the late 1980s, to the point that traditional measures of predictive ability (that evaluate
the performance of the model on average over the out-of-sample period) would overstate
the recent predictive ability of the UIRP model. Rossi and Sekhposyan (2011) apply new
methodologies to better understand why the economic models’performance is poor. They
…nd that lack of predictive content is the major explanation for the lack of short-term
forecasting ability of the economic models, whereas instabilities play a role especially for
medium term (one-year ahead) forecasts.

What we learn from this section is that the choice of the forecast evaluation method matters
a lot. In particular, we note the following facts. The choice of the benchmark model is very
important: choosing an inappropriate benchmark model overstates the empirical evidence
in favor of the economic model’s predictive ability; the random walk (without drift) is the
toughest benchmark to beat and should be the one used in practice. The forecast horizon
also matters a lot: typically, monetary fundamentals do not have short-horizon predictive
ability (1 month or 1 quarter), although the monetary fundamentals in an ECM model
may forecast better than a random walk at long horizons (3-4 years). In general, both the
evaluation method as well as the forecast sample may potentially and substantially matter.
For example, UIRP may outperform the random walk if certain tests are used (CW rather
than DMW); in addition, its performance is unstable over time and worsened in the last
decades. Rogo¤ and Stavrakeva (2008) …nd that the out-of-sample forecasting ability of the
panel ECM model is not robust either to the choice of the forecast window nor to the choice
of the out-of-sample forecast period. To conclude, Panel D in Table 2 summarizes whether
the predictors in the selected papers we consider are successful in forecasting exchange rates.
71
See Rossi (2011) for an overview of forecast tests in the presence of instabilities.
38

7 Empirical Analysis
To summarize, there is substantial disagreement in the literature regarding whether exchange
rates are forecastable. The literature clearly reached a consensus on several "negative"
stylized facts, namely the fact that the monetary and PPP fundamentals have no predictive
ability at short horizons. There is also agreement on the fact that non-linear models are the
least successful models. Finally, there is also agreement on the fact that, should monetary
models have any predictive ability at long horizons, it only appears in single-equation ECM
and panel ECM model speci…cations. On the other hand, there is substantial disagreement
on whether monetary fundamentals do have predictive ability at long horizons and whether
UIRP has predictive ability at short horizons. There is some evidence in favor of BMA
models, although not strong. The only "positive" …nding the literature reached a consensus
on seems the fact that Taylor-rules and net foreign assets models have signi…cant forecasting
ability at short horizons.
In this last section we undertake an empirical analysis to illustrate several of the di¤erent
approaches, models and methods used in the literature, and revisit the empirical …ndings
using the most up-to-date sample, methods and fundamentals we have available. We focus
on univariate linear models with either traditional or newly proposed fundamentals, the
single-equation ECM and panel ECM monetary models as well as the BMA. Since non-linear
models …t well in-sample, but produce poor out-of-sample forecasts, we do not consider them
in our empirical exercise; similarly, the most recent studies report lack of empirical support
for time-varying parameter models, so we will not consider them either.
Given that the literature di¤ers on the data transformations, we will focus on a unifying
framework where the only data transformation is seasonal adjustment, which we perform
consistently for all series. We use several of the methods reviewed in Section 6 in the
empirical analysis.72
7.1 Data Description

We collect data on exchange rates (relative to the US) and several economic fundamentals
for a variety of countries: Australia, Austria, Belgium, Canada, Denmark, Finland, France,
Germany, Greece, Ireland, Italy, Japan, New Zealand, Portugal, Spain, Sweden, Switzerland,
72
Data limitations prevent us from studying the importance of real-time versus revised data for the coun-
tries that we consider.
39

UK and the US.73 For all countries, we collect monthly and quarterly data on overnight in-
terest rates, 3-month Treasury Bills, 5-year Treasury Bonds, GDP/ industrial production,
CPI and the money stock. We also collect data on the current account, the trade balance,
public debt and government de…cit/surplus – the latter data are available only at the an-
nual or quarterly frequency. Data are from the IMF database via Datastream as well as
Philip Lane’s website (http://www.philiplane.org/EWN.html). Since countries’geographi-
cal de…nitions have changed over time (for example, after the start of the euro), the sample
size di¤ers across countries. Also, slightly di¤erent fundamentals are available depending on
the country (for example, the exact de…nition of money supply available might di¤er across
countries, and the sample sizes for our measure of output for the same country di¤er depend-
ing on the frequency since GDP is only available at the quarterly frequency, and industrial
production is available at the monthly frequency). We selected the most homogeneous def-
initions we could …nd, while striving to collect data that have the longest available sample.
Table 3 provides information on the available samples. A not-for-publication Appendix pro-
vides additional details on the data, as well as their mnemonics and sources. The series are
originally not seasonally adjusted, and seasonal adjustment was performed using one-sided
moving averages with backward, equal weights.74
7.2 Methodology
We present empirical results using a few of the most used predictors and successful method-
ologies. Regarding the predictors, we consider interest rate di¤erentials, price di¤erentials,
money and output di¤erentials, Taylor rule fundamentals and measures of external imbalance
such as the current account and the trade balance.75
Regarding the models, …rst we consider the performance of selected traditional single-
equation linear models (UIRP, eq. 1; PPP, eq. 2, and monetary) as well as Taylor-rule
fundamentals (eq. 8) and portfolio balance models (eq. 6).76 The monetary model is
73
While we collected data also for Argentina, Bolivia, Brazil, Chile, Colombia, Costa Rica, Hong Kong,
Mexico, Peru, Portugal, Russia, Singapore, Taiwan, Thailand and Turkey, the sample sizes of either the
exchange rate or the fundamentals were severely limited, or potentially severely a¤ected by measurement
error. These countries were discarded from the analysis.
74
For quarterly data, the …lter is 1=3+1=3L+1=3L2 ; for monthly data it is (1=12)+(1=12) L+:::+(1=12) L11 .
Empirical results based on seasonally unadjusted data are qualitatively similar.
75
The analysis will focus on monthly data when possible, given that both monthly and quarterly data
have the same span but the former provides larger sample sizes.
76
Unfortunately, net foreign assets data are available only at the annual frequency, so an analysis with the
40

considered either as in eq. (3) with variables in …rst di¤erences, or in the ECM form, eq.
(12).77 We will then turn to two representative multivariate models (the BMA and the panel
VECM models). We will compare the models’forecasting performance at short (one month
or one quarter) and long (four years) horizons with that of a random walk without drift
using RMSFE comparisons. The models’ parameters are recursively re-estimated with a
rolling window whose size is equal to half of the total sample size in the baseline analysis,
although we investigate the consequences of varying both the size of the window and the
forecast sample. Since the literature review highlights the importance of the forecast sample,
we formally investigate the robustness (or lack thereof) of the empirical …ndings to both the
choice of the out-of-sample forecast period and the choice of the estimation window size
by using the tests proposed by Giacomini and Rossi (2010) and Inoue and Rossi (2012),
respectively.78
We will compare models’ performance using both in-sample and out-of-sample tests.
As discussed in the previous section, out-of-sample forecast tests may have di¤erent power
properties, so in the empirical application, we shed some light on this issue by considering
both the traditional DMW and the CW tests.
7.3 Empirical Evidence

Results are reported in Tables 4 and 5. More in detail, for each model we consider, the …rst
column reports the country whose nominal exchange rate growth we are forecasting (relative
to the US dollar). The second column (labeled "GC") reports p-values of Rossi’s (2005b)
Granger-causality test robust to instabilities.79 The third column reports the ratio of the
root mean squared forecast error of the model relative to that of the random walk without
drift (labeled "RMSFER"): values smaller than unity denote that the model forecasts better
than the random walk benchmark. Finally, the last two columns assess the signi…cance of
the test for out-of-sample forecasting ability: the column labeled "DMW" reports p-values
of the Diebold and Mariano (1995) and West (1996) test and the column labeled "CW"
techniques currently used in this paper is not possible due to the small sample sizes.
77
We report results for calibrated cointegration parameters: = 1, as in Mark (1995) –see the discussion
in Section 4.
78
Note that since the total sample size varies by country, the size of the rolling window as well as the
out-of-sample forecast sample vary by country, too.
79
Detailed empirical results based on traditional Granger-causality tests are reported in the not-for-
publication appendix and their …ndings summarized in this section.
41

reports p-values of the Clark and West (2006) test.80 The benchmark model in the latter
two tests is the random walk without drift. "h" denotes the forecast horizon. All tests are
implemented with Newey and West’s (1987) heteroskedasticity and serial correlation robust
covariance matrix.81 Note that p-values lower than, say, 0.05 denote signi…cance at the 5%
level, which is the de…nition of signi…cance we adopt in what follows when discussing the
results. In all the tables, "n.a." denotes cases in which the sample size was too small and
the test statistic is not reported.
Traditional Predictors’ Performance in Monthly Data. Table 4, Panels A-D
report the performance of the traditional predictors. The panels show some in-sample
predictive ability at short (one-month-ahead) horizons for UIRP and the monetary ECM
models,82 but very limited at long (four-year-ahead) horizons for any of the models.83 The
out-of-sample predictive ability is almost inexistent: typically, models’ MSFEs are larger
than the random walk’s,84 and neither DMW nor CW …nd signi…cant predictive ability.
Why is it that our results for the monetary ECM model are so di¤erent from those in the
literature? We investigate this issue at the end of this section.
Alternative Predictors’Performance. Panel E in Table 4 shows empirical evidence
in favor of in-sample predictability of Taylor rule fundamentals for several countries.85 The
CW test …nds empirical evidence in favor of the Taylor-rule model only at short horizons: it
is strongly signi…cant for six countries and marginally for two others at the one-month-ahead
horizon, and for four of the countries at the four-year horizon. Among the fundamentals that
we consider, then, they are among the most successful out-of-sample at short horizons. Im-
portantly, note however that the DMW test does not …nd predictive ability. The discrepancy
between the CW and DMW tests emphasizes that the way we treat parameter estimation
error matters: if we compare models and correct inference for the fact that the Taylor model
80
While the DMW test is typically used for non-nested models’ forecast comparisons, in this article we
interpret its critical values following Giacomini and White (2006), which show that the DMW test can be
used to compare forecasts of nested models provided a rolling window is used for estimation.
81
The truncation parameter is T 1=4 , where T is the available sample size.
82
In fact, seven countries are signi…cant for UIRP, three for PPP, one for the monetary and nine for the
monetary ECM models. These results show that, for example, interest rates have been capable, at some
point in time, to predict one-month-ahead exchange rates. According to the traditional Granger-causality
test, reported in the not-for-publication appendix, traditional fundamentals have more limited predictive
ability, signi…cant only for a couple of countries for each model/predictor.
83
PPP is signi…cant for four countries, the other models/predictors perform much worse.
84
MSFER is bigger than or equal to one.
85
We consider the symmetric, homogeneous Taylor rule, eq. (8).
42

estimates more parameters than the random walk, we conclude that the Taylor model, when
evaluated "in population", has better predictive ability; however, if we evaluate its forecasts
"at the actual estimated parameter values", their forecasting ability is not superior to that
of the random walk.
Panel F in Table 4 reports results for the model that includes real interest rates, the trade
balance and the current account as predictors (at the quarterly frequency). Interestingly,
among the models we consider, this is one with the strongest in-sample predictive ability;
however, its out-of-sample performance is extremely poor (possibly due to the large number
of parameters to estimate).
Multivariate Models. Results for multivariate models are reported in Table 5. Panel A
in Table 5 shows that the monetary panel model displays some in-sample predictive ability
(at some point in time) and almost inexistent out-of-sample forecasting ability at short
horizons; interestingly, however, there is some evidence of out-of-sample forecasting ability
at long horizons for four countries. Again, note the very di¤erent results for CW and DMW:
once more, the latter never …nds predictive ability. Panel B in Table 5 shows instead that
the BMA is not signi…cantly better than the random walk benchmark in forecasting out-of-
sample at any horizon and for any test statistics.
Instabilities in Predictive Performance. Our results are broadly consistent with
those in the literature, although in some cases they di¤er. For example, we …nd less out-
of-sample forecasting ability at long horizons than other papers in the literature for the
monetary model in error-correction form (ECM). Two of the possible reasons why our results
may di¤er from those reported by other papers are that either the out-of-sample period is
di¤erent, or the window size is di¤erent, or both. Let us consider each of these two e¤ects
separately. We report results based on the Giacomini and Rossi (2010) and the Inoue and
Rossi (2012) tests in Figures 1 to 6.
Figures 1, 2, 5, and 6 plot Giacomini and Rossi’s (2010) Fluctuation test statistic (solid
line) along with its critical value (dotted line) for selected countries, models and forecast
horizons.86 The x-axis denotes time. The Fluctuation test examines how the forecast per-
formance changed over time: whether the out-of-sample predictive ability of the traditional
predictors ever showed up in the data, and, if so, when. For example, in Figure 1 the Fluc-
tuation test (solid line) for Canada is never above its critical value (dotted line), thus the
predictor never displays signi…cant forecasting ability; on the other hand, the Fluctuation
test does detect predictive ability for France in the late 2000s, Japan in mid-2000 and the
86
The Fluctuation test is implemented with m = 1=3 in Giacomini and Rossi’s (2010) notation.
43

UK in 2009. Overall, Figure 1 shows that monetary fundamentals have occasional and very
short-lived predictive ability at the one-month horizon for some countries at some point in
time. Figure 2 shows similar, occasional predictive ability at long horizons.87
A second reason why the results may di¤er is because of the use of a di¤erent estimation
window size.88 We study the behavior of the predictive ability as a function of the window
size in Figures 3 and 4. The …gures plot Inoue and Rossi’s (2012) test statistic (solid line)
along with its critical value (horizontal line) for selected countries, models and forecast
horizons. The x-axis denotes the window size, R.89 For example, Figure 3 reports the DMW
test calculated in rolling windows over the out-of-sample period for various window sizes R,
reported on the x-axis. When the DMW test (solid line) is above Inoue and Rossi’s (2012)
critical values (dotted line), this signals predictive ability for that window size. Overall,
Figures 3 and 4 show that indeed the window size strongly a¤ects predictability for some
countries.90
However, note that the out-of-sample period changes as we change the estimation window
size: a larger estimation window size implies that a larger proportion of the sample is used
for estimation and a smaller proportion is used for out-of-sample forecast evaluation. To
analyze how the choice of the estimation window size and the out-of-sample period interact
with each other, Figures 5 and 6 report the Fluctuation test for various window sizes for
Germany and Japan, respectively. Very interesting …ndings emerge from the …gures. Figure
5 shows that only small estimation window sizes can detect forecasting ability of monetary
fundamentals in Germany, and their predictive power is concentrated in the late 1980s;
Figure 6 shows instead that only large estimation window sizes can detect predictability of
monetary fundamentals in Japan, and that the predictability emerges in the late 2000s.
Clearly, for some window sizes, it is possible to …nd some evidence of predictive ability
for monetary fundamentals, thus con…rming previous results that have been reported in the
literature for selected window sizes and/or out-of-sample evaluation periods. At the same
87
See Giacomini and Rossi (2010) for time variation in Taylor-rule fundamentals.
88
In our exercise, we choose a rolling window with a size equal to half of the total sample size: while
the fact that we select a window size equal to half of the total sample size is similar to what other papers
have done, still our total sample is longer, and therefore our results may di¤er from those in the literature;
furthermore, the sample is di¤erent too.
89
We implement the sup-type version of Inoue and Rossi’s (2012) test.
90
Figure 3 shows predictability at the one-month-ahead horizon for Switzerland, Japan and the UK: in
the former, predictability is enhanced by a large window size, in the latter two it is the opposite. Figure 4
shows similar results for 48–month-ahead forecasts for Switzerland and Japan.
44

time, our results highlight the lack of robustness of these analyses to small changes in the
procedures for forecast estimation and evaluation.
Summary
To summarize our results, most traditional predictors show in-sample forecasting ability,
but none of them shows strong out-of-sample forecasting ability across all countries and tests.
Turning to the most recent Taylor-rule predictors, they are among the most successful fun-
damentals out-of-sample at short horizons. Exploiting information on fundamentals using
panels does not seem to help at short horizons, although it provides forecast improvements
over the random walk at long horizons for a few countries. Figures 7 and 8 summarize the
empirical …ndings. Figure 7 is a scatterplot of the p-values of Clark and West’s (2006) test
at short horizons (reported on the x-axis) and at long horizons (reported on the y-axis)
for several predictors, denoted by di¤erent markers. Each point in the …gure corresponds
to a country/ predictor/ model combination. P-values smaller than 0.05 denote statistical
signi…cance at the 5% level. The …gure shows that, for several countries, Taylor-rule fun-
damentals are concentrated in the left, mid-to-upper corner of the …gure; therefore they are
signi…cant predictors at short horizons but not at long horizons. Conversely, the monetary
panel model shows up on the right side of the …gure; in some cases in the lower bottom
and in several other cases on the top, right corner. Therefore, the monetary panel model is
signi…cant at long horizons (but not at short horizons) for some countries, and not signi…-
cant at any horizon for several other countries. Figure 8 does the same for Rossi’s (2005b)
robust Granger-causality test. It shows that Taylor-rule and monetary panel models per-
form the best in-sample (the former at both horizons, the latter at short horizons only);
however, comparing Figures 7 and 8 clearly shows that only the in-sample predictive ability
of Taylor-rules survives out-of-sample.
Only when we allow the predictability to be varying over time using the Fluctuation test,
we see a few episodes when the economic predictors’forecasting ability was stronger than that
of the random walk benchmark even in traditional predictors, although they remain sporadic
and short-lived in most countries. So, forecasting ability, when present, is an occasional and
short-lived phenomenon.
8 Conclusions
To conclude, our analysis of the literature review has uncovered several stylized facts, which
lead to …ve main conclusions.
45

First, a consensus in the literature is that the Taylor-rule and net foreign assets funda-
mentals have more out-of-sample predictive content than traditional fundamentals (such as
interest rate, in‡ation, output and money di¤erentials); in fact, monetary fundamentals at
very long horizons and interest rate di¤erentials at short horizons display forecasting ability
according to some papers but not others. However, the disagreement in the literature is in
the extent to which the former can explain the Meese and Rogo¤ puzzle.
Second, overall, among the model speci…cations considered in the literature, the most
successful are linear speci…cations. Among them, the single-equation ECM and the panel
ECM models are the most successful at long horizons, although there is disagreement among
researchers about the degree of robustness of the results. Typically, but not always, for
single-equation linear models the predictor choice matters more than whether the researcher
uses contemporaneous, realized or lagged fundamentals. Among the multivariate models,
the most successful speci…cation is the panel ECM.
Third, data transformations (such as de-trending, …ltering and seasonal adjustment) may
substantially a¤ect predictive ability, and may explain some di¤erences in results across
studies. For example, the forecasting ability of the monetary model at long horizons is much
weaker or completely disappears after estimating the cointegration parameters. Another
important factor that, for some fundamentals, may a¤ect predictive ability is using realized
or real-time data.91 For a given model and predictor, predictive ability seems also to depend
on the choice of the country; on the other hand, with few exceptions, the frequency of the
data and whether the realized or the forecasted fundamental is used do not seem to a¤ect
predictive ability.
Fourth, the choice of the benchmark, horizon, sample period and forecast evaluation
method matters a lot. In particular, choosing an inappropriate benchmark model may over-
state the empirical evidence in favor of the economic model’s predictive ability: the random
walk without drift is the toughest benchmark to beat. Empirical results depend on the
forecast horizon,92 the choice of the evaluation method,93 and the forecast sample, as the
performance of predictors is typically unstable over time.94
91
This is a concern for monetary fundamentals and less of a concern for Taylor-rule fundamentals.
92
For example, the literature agrees that monetary fundamentals do not have predictive ability over short
horizons, but disagrees on whether they might have over long horizons.
93
For example, whether interest rate di¤erentials outperform the random walk may depend on the test
that is used in practice.
94
For example, the out-of-sample forecasting ability of the panel ECM model is not robust to either the
choice of the forecast window or the forecast period.
46

Fifth, on the one hand, the empirical analysis con…rms several of the …ndings in the
literature: several predictors display in-sample predictive ability for future exchange rates;
however, only few predictors display out-of-sample forecasting ability: Taylor-rules at short
horizons for several countries, and monetary panel models at long horizons for some countries.
On the other hand, our literature review and empirical evidence suggest slightly less out-of-
sample predictive ability in favor of some of the models and predictors used in the literature.95
The reason is that there are substantial instabilities in the models’forecasting performance:
the predictive power of fundamentals varies not only across countries, models and predictors,
but also over time periods; it may appear in some periods and disappear in others.
Overall, although some predictors (Taylor-rule fundamentals and net foreign assets) do
exhibit some predictive ability at short horizons, and others (monetary fundamentals, espe-
cially in panel models) reveal some predictive ability at long horizons, none of the predictors,
models, or tests systematically …nd empirical support for superior exchange rate forecasting
ability of a predictor for all models, countries and time periods: typically, when predictabil-
ity appears, it does so occasionally for some countries and for short periods of time. Thus,
Meese and Rogo¤’s (1983a,b) …nding does not seem to be entirely and convincingly over-
turned. We also …nd another interesting puzzle: the predictive ability of the fundamentals is
time-varying and occasional, yet existing time-varying parameter models are not successful
in capturing it. Our …ndings lead to new challenges: why does the predictability of exchange
rate models change over time? Is it possible to design ways to exploit instabilities and im-
prove exchange rates’forecasts? An answer to these questions would require an answer to
another important question: why are exchange rates poorly forecasted by economic models?
In their papers, Meese and Rogo¤ (1983a,b) conjectured that sampling error, model mis-
speci…cation and instabilities could potentially explain the poor forecasting performance of
the economic models. Insights on the latter may be particularly relevant to understand and
resolve the Meese and Rogo¤ puzzle.
95
For example, monetary fundamentals at long horizons.
47

9 References
Abhyankar, A., L. Sarno and G. Valente (2005), “Exchange Rates and Fundamentals: Evi-
dence on the Economic Value of Predictability,” Journal of International Economics 66(2)
325-348.
Adrian, T., E. Etula, and H.S. Shin (2011), “Risk Appetite and Exchange Rates,”mimeo.
Alquist, R. and M.D. Chinn (2008), “Conventional and Unconventional Approaches to
Exchange Rate Modelling and Assessment,”International Journal of Finance and Economics
13, 2-13.
Amano, R.A. and S. van Norden (1993), “A Forecasting Equation for the Canada-US Dol-
lar Exchange Rate,” in: The Exchange Rate and the Economy, Proceedings of a conference
held at the Bank of Canada.
Amano, R.A. and S. van Norden (1995), “Terms of Trade and Real Exchange Rates: the
Canadian Evidence,”Journal of International Money and Finance 14(1), 83-104.
Amano, R.A. and S. van Norden (1998a), “Oil prices and the Rise and Fall of the US
Real Exchange Rate,”Journal of International Money and Finance 17(2), 299-316.
Amano, R.A. and S. van Norden (1998b), “Exchange Rates and Oil Prices,” Review of
International Economics 6(4), 683-94.
Andersen, T., T. Bollerslev, F.X. Diebold and C. Vega (2003), “Micro E¤ects of Macro
Announcements: Real-Time Price Discovery in Foreign Exchange,” American Economic
Review 93(1), 38-62.
Bacchetta, P., E. van Wincoop and T. Beutler (2010), “Can Parameter Instability Explain
the Meese-Rogo¤ Puzzle?,”in: L. Reichlin and K. West (eds.), NBER International Seminar
on Macroeconomics, University of Chicago Press.
Backus, D., S. Foresi and C. Telmer (2001), “A¢ ne Term Structure Models and the
Forward Premium Anomaly,”Journal of Finance 56, 279–304.
Baillie, R.T. and T. Bollerslev (1989), “Common Stochastic Trends in a System of Ex-
change Rates,”Journal of Finance 44, 167–181.
Baillie, R.T. and T. Bollerslev (1994), “Cointegration, Fractional Cointegration and Ex-
change Rate Dynamics,”Journal of Finance 49, 737–745.
Baillie, R.T. and T. Bollerslev (2000), “The Forward Premium Anomaly is Not as Bad
as You Think,”Journal of International Money and Finance 19(4), 471-488.
Bailliu, J. and M.R. King (2005), “What Drives Movements in Exchange Rates?”, Bank
of Canada Review, Autumn.
48

Balassa, B. (1964), “The Purchasing Power Parity Doctrine: A Reappraisal,”Journal of
Political Economy 72, 584-596.
Bekaert, B. and R. Hodrick (2001), “Expectations Hypotheses Tests,”Journal of Finance
56(4), 1357-1394.
Berben, R.B. and D.J. van Dijk (1998), “Does the Absence of Cointegration Explain the
Typical Findings in Long Horizon Regressions?,”Erasmus University Econometrics Institute,
Report 9814.
Berkowitz, J. and L. Giorgianni (2001), “Long-Horizon Exchange Rate Predictability?,”
Review of Economics and Statistics 83(1), 81-91.
Canova, F. (1993), “Modelling and Forecasting Exchange Rates with a Bayesian Time-
Varying Coe¢ cient Model,”Journal of Economic Dynamics and Control 17(1/2), 233-61.
Carriero, A., G. Kapetanios and M. Marcellino (2009), “Forecasting Exchange Rates with
a Large Bayesian VAR,”International Journal of Forecasting 25, 400-417.
Cassel, G. (1918), “Abnormal Deviations in International Exchanges,”Economic Journal
28,413–15.
Cerra, V. and S.C. Saxena (2010), “The Monetary Model Strikes Back: Evidence from
the World,”Journal of International Economics 81, 184-196.
Chen Y.C. and K.S. Rogo¤ (2003), “Commodity Currencies,” Journal of International
Economics 60, 133-169.
Chen, Y.C., K.S. Rogo¤ and B. Rossi (2010), “Can Exchange Rates Forecast Commodity
Prices?,”Quarterly Journal of Economics 125(3), 1145-1194.
Chen, Y.C. and B.K.P. Tsang (2011), “What Does the Yield Curve Tell Us About Ex-
change Rate Predictability?,”Review of Economics and Statistics, forthcoming.
Cheung, Y.W., M.D. Chinn and A.G. Pascual (2005), “Empirical Exchange Rate Models
of the Nineties: Are Any Fit to Survive?,”Journal of International Money and Finance 24,
1150-1175.
Cheung, Y.W. and U.G. Erlandsson (2005), “Exchange Rates and Markov Switching
Dynamics,”Journal of Business and Economic Statistics 23, 314-320.
Cheung, Y. and K. Lai (2000), “On the Purchasing Power Parity Puzzle,” Journal of
International Economics 52, 321–330.
Chinn, M. (1991), “Some Linear and Nonlinear Thoughts on Exchange Rates,” Journal
of International Money and Finance 10, 214-230.
Chinn, M. (2008), “Nonlinearities, Business Cycles and Exchange Rates,” Economic
Notes 37(3), 219-239.
49

Chinn, M. (2009), “Comment on: Can Parameter Instability Explain the Meese-Rogo¤
Puzzle?,” in: L. Reichlin and K. West (eds.), NBER International Seminar on Macroeco-
nomics, University of Chicago Press.
Chinn, M. (2011), “Macro Approaches to Foreign Exchange Determination,”La Follette
School Working Paper 2011-013.
Chinn, M.D. and R. Meese (1995), “Banking on Currency Forecasts: Is Change in Money
Predictable?,”Journal of International Economics 38 (1-2), 161-78.
Chinn, M.D. and G. Meredith (2004), “Monetary Policy and Long-Horizon Uncovered
Interest Parity,”IMF Sta¤ Papers 51(3).
Chinn, M.D. and M.J. Moore (2012), “Order Flow and the Monetary Model of Exchange
Rates: Evidence from a Novel Data Set,”Journal of Money, Credit and Banking, forthcom-
ing.
Clarida, R., J. Gali and M. Gertler (1998), “Monetary Rules in Practice: Some Interna-
tional Evidence,”European Economic Review 42, 1033-1067.
Clarida, R.H., L. Sarno, M.P. Taylor and G. Valente (2003), “The Out-of-sample Suc-
cess of Term Structure Models as Exchange Rate Predictors: A Step Beyond,” Journal of
International Economics 60(1), 61-83.
Clarida, R.H. and M.P. Taylor (1997), “The Term Structure Of Forward Exchange Pre-
miums And The Forecastability Of Spot Exchange Rates: Correcting The Errors,” Review
of Economics and Statistics 79(3), 353-361.
Clark, T.E. and M.W. McCracken (2001), “Tests of Equal Forecast Accuracy and En-
compassing for Nested Models,”Journal of Econometrics 105, 85-110.
Clark, T.E. and K.D. West (2006), “Using Out-of-Sample Mean Squared Prediction
Errors to Test the Martingale Di¤erence Hypothesis,” Journal of Econometrics 135, 155-
186.
Clark, T.E. and K.D. West (2007), “Approximately Normal Tests for Equal Predictive
Accuracy in Nested Models,”Journal of Econometrics 138, 291-311.
Coletti, D. and S. Murchison (2002), “Models in Policy-Making,”Bank of Canada Review.
Corradi, V. and N. Swanson (2006), “Predictive Density Evaluation,”in: Elliott, G., C.
Granger and A. Timmermann (eds.), Handbook of Economic Forecasting Vol. 1, Elsevier-
North Holland, 100-134.
Croushore, D. and T. Stark (2001), “A Real-time Data Set for Macroeconomists,”Journal
of Econometrics 105(1), 111-130.
Croushore, D. and T. Stark (2003), “A Real-Time Data Set for Macroeconomists: Does
50

the Data Vintage Matter?,”Review of Economics and Statistics 85(3), 605-617.
Della Corte, P., L. Sarno and G. Sestieri (2010), “The Predictive Information Content
of External Imbalances for Exchange Rate Returns: How Much Is It Worth?,” Review of
Economics and Statistics, forthcoming.
Diebold, F.X., J. Hahn and A. Tay (1999), “Multivariate Density Forecast Evaluation and
Calibration in Financial Risk Management: High-Frequency Returns on Foreign Exchange,”
Review of Economics and Statistics 81, 661-673.
Diebold, F., J. Gardeazabal and K. Yilmaz (1994), “On Cointegration and Exchange
Rate Dynamics,”Journal of Finance 49 (2), 727–735.
Diebold, F.X. and R.S. Mariano (1995), “Comparing Predictive Accuracy,” Journal of
Business and Economic Statistics 13, 253-263.
Diebold, F.X. and J.A. Nason (1990), “Nonparametric Exchange Rate Prediction?,”
Journal of International Economics 28(3-4), 315-332.
Diebold, F.X. (2012), “Comparing Predictive Accuracy, Twenty Years Later: A Personal
Perspective on the Use and Abuse of Diebold-Mariano Tests," mimeo, U. of Pennsylvania.
Dimand, R.W. (1999), “Irving Fisher and the Fisher Relation: Setting the Record
Straight,”The Canadian Journal of Economics 32 (3), 744-750.
Edge, R.M., M.T. Kiley and J.P. Laforte (2010), “A Comparison of Forecast Performance
between Federal Reserve Sta¤ Forecasts, Simple Reduced Form Models and a DSGE Model,”
Journal of Applied Econometrics 25(4), 720–754.
Ehrmann, M. and M. Fratzscher (2005), “Exchange Rates and Fundamentals: New Evi-
dence from Real-time Data,”Journal of International Money and Finance 24, 317-341.
Elliott, G. and A. Timmermann (2008), “Economic Forecasting,” Journal of Economic
Literature 46(1): 3–56.
Engel, C.M. (1994), “Can the Markov Switching Model Forecast Exchange Rates?,”
Journal of International Economics 36, 151–165.
Engel, C.M. and J.D. Hamilton (1990), “Long Swings in the Exchange Rate: Are They
in the Data and Do Markets Know It?,”American Economic Review 80, 689-713.
Engel, C., N.C. Mark and K.D. West (2007), “Exchange Rate Models Are Not as Bad as
You Think,”in: NBER Macroeconomics Annual 2007, Acemoglu, D., Rogo¤, K., Woodford,
M., eds., University of Chicago Press, 381-441.
Engel, C., N.C. Mark and K.D. West (2009), “Factor Model Forecasts of Exchange
Rates,”mimeo.
Engel, C. and K.D. West (2004), “Accounting for Exchange-Rate Variability in Present-
51

Value Models When the Discount Factor is Near 1,” American Economic Review 94(2),
119-125.
Engel, C. and K.D. West (2005), “Exchange Rate and Fundamentals,”Journal of Political
Economy 113, 485-517.
Engel, C. and K.D. West (2006), “Taylor Rules and the Deutschemark-Dollar Real Ex-
change Rate,”Journal of Money, Credit and Banking 38, 1175-1194.
Fama, E. (1984), “Forward and Spot Exchange Rates,”Journal of Monetary Economics
14, 319–338.
Faust, J., J.H. Rogers, S.Y.B. Wang and J.H. Wright (2007), “The High-Frequency Re-
sponse of Exchange Rates and Interest Rates to Macroeconomic Announcements,” Journal
of Monetary Economics 54, 1051-1068.
Faust, J., J.H. Rogers and J.H. Wright (2003), “Exchange Rate Forecasting: the Errors
We’ve Really Made,” Journal of International Economics 60(1), 35-59.
Ferraro, D., K.S. Rogo¤ and B. Rossi (2011), “Can Oil Prices Forecast Exchange Rates?,”
ERID Working Paper No. 95, Duke University.
Fisher, I. (1896), Appreciation and Interest, New York: Macmillan for the American
Economic Association; reprinted in: The Works of Irving Fisher, Vol. 1, W.J. Barber (ed.),
London: Pickering & Chatto.
Frankel, J.A. (1982), “The Mystery of the Multiplying Marks: A Modi…cation of the
Monetary Model,”Review of Economics and Statistics 64, 515-519.
Fratzscher, M. (2009), “What Explains Global Exchange Rate Movements During the
Financial Crisis,”Journal of International Money and Finance 28, 1390-1407.
Frankel, J.A. and A.K. Rose (1995), “Empirical Research on Nominal Exchange Rates,”
in: Grossman, G. and K.S. Rogo¤ (eds.), Handbook of International Economics Vol. 3,
Elsevier-North Holland.
Frenkel, J.A. (1976), “A Monetary Approach to the Exchange Rate: Doctrinal Aspects
and Empirical Evidence,”Scandinavian Journal of Economics 78, 200-224.
Froot, K. and K. Rogo¤ (1995), “Perspectives on PPP and Long-run Real Exchange
Rates,”in: Grossman, G. and K.S. Rogo¤ (eds.), Handbook of International Economics Vol.
3, Elsevier-North Holland.
Froot, K. A. and R. H. Thaler (1990), “Anomalies: Foreign Exchange,”Journal of Eco-
nomic Perspectives 4, 179–192.
Giacomini, R. and B. Rossi (2010), “Forecast Comparisons in Unstable Environments,”
Journal of Applied Econometrics 25(4), 595-620.
52

Giacomini, R. and H. White (2006), “Tests of Conditional Predictive Ability,” Econo-
metrica 74(6), 1545-1578,
Giannone, D. (2009), “Comment on: Can Parameter Instability Explain the Meese-Rogo¤
Puzzle?,” in: L. Reichlin and K. West (eds.), NBER International Seminar on Macroeco-
nomics, University of Chicago Press.
Gospodinov, N. (2004), “Asymptotic Con…dence Intervals for Impulse Responses for
Near-Integrated Processes: An Application to Purchasing Power Parity,”Econometrics Jour-
nal 7(2), 505-527.
Gourinchas, P.O. and H. Rey (2007), “International Financial Adjustment,” Journal of
Political Economy 115(4), 665-703.
Greenspan, A. (1994), “A Discussion of the Development of Central Banking,”in: Capie,
F., C. Goodhart, S. Fischer and N. Schnadt (eds.), The Future of Central Banking, Cam-
bridge: Cambridge University Press.
Groen, J.J. (1999), “Long Horizon Predictability of Exchange Rates: is it for Real?,”
Empirical Economics 24, 451–469.
Groen, J.J. (2000), “The Monetary Exchange Rate Model as a Long-run Phenomenon,”
Journal of International Economics 52, 299–319.
Groen, J.J. (2002), “Cointegration and the Monetary Exchange Rate Model Revisited,”
Oxford Bulletin of Economics and Statistics 64(4), 361-380.
Groen, J.J. (2005), “Exchange Rate Predictability and Monetary Fundamentals in a
Small Multi-Country Panel,”Journal of Money, Credit and Banking 37, 495-516.
Hooper, P. and J.E. Morton (1982), “Fluctuations in the Dollar: A Model of Nominal
and Real Exchange Rate Determination,” Journal of International Money and Finance 1,
39-56.
Husted, S. and R. MacDonald (1998), “Monetary-based Models of the Exchange Rate:
a Panel Perspective,” Journal of International Financial Markets, Institutions and Money
8(1), 1-19.
Imbs, J., H. Mumtaz, M. Ravn and H. Rey (2005), “PPP Strikes Back: Aggregation and
the Real Exchange Rate,”Quarterly Journal of Economics 120(1), 1-43.
Inoue, A. and B. Rossi (2008), “Monitoring and Forecasting Currency Crises,” Journal
of Money, Credit and Banking 40(2-3), 523-534.
Inoue, A. and B. Rossi (2012), “Out-of-Sample Forecast Tests Robust to the Choice of
Window Size,” Journal of Business and Economic Statistics 30(3), 432-453.
Ito, T. (1990), “Foreign Exchange Rate Expectations: Micro Survey Data,” American
53

Economic Review, 80, 434-449.
Keynes, J.M. (1923), A Tract on Monetary Reform, London: Macmillan. Reprinted in:
D.E. Moggridge and E.A.G. Robinson (eds.), Collected Writings of John Maynard Keynes
Vol. 4, Cambridge University Press for the Royal Economic Society, 1971.
Kilian, L. (1999), “Exchange Rates and Monetary Fundamentals: What Do We Learn
From Long-Horizon Regressions?,”Journal of Applied Econometrics 14(5).
Kilian, L. and M.P. Taylor (2003), “Why Is It so Di¢ cult to Beat the Random Walk
Forecast of Exchange Rates?,”Journal of International Economics 60(1), 85-107.
Kilian, L. and T. Zha (2002), “Quantifying the Uncertainty About the Half-Life of De-
viations From PPP,”Journal of Applied Econometrics 17, 107–125.
Koenig, E., S. Dolmas and J. Piger (2003), “The Use and Abuse of “Real-time”Data in
Economic Forecasting,”Review of Economics and Statistics 85(3), 618-628.
Lane P. and G.M. Milesi-Ferretti (2007), “The External Wealth of Nations II; Revised and
Extended Estimates of Foreign Assets and Liabilities 1970–2004,” Journal of International
Economics 73(2), 223-250.
Lewis, K. (1995), “Puzzles in International Finance,” in: Grossman, G.M. and K.S.
Rogo¤ (eds.), Handbook of International Economics Vol. 3, Elsevier-North Holland, 1913-
1949.
Lopez, C. C. Murray, and D. Papell (2005), “State-of-the-Art Unit Root Tests and the
PPP Puzzle,”Journal of Money, Credit and Banking 37, 361-69.
MacDonald, R. and M.P. Taylor (1993), “The Monetary Approach to the Exchange Rate:
Rational Expectations, Long-run Equilibrium, and Forecasting,”IMF Sta¤ Papers 40.
MacDonald, R. and M.P. Taylor (1994), “The Monetary Model of the Exchange Rate:
Long-run Relationships, Short-run Dynamics, and How to Beat a Random Walk,” Journal
of International Money and Finance 13, 276–290.
Marcellino, M., J.H. Stock and M.W. Watson (2006), “A Comparison of Direct and
Iterated AR Methods for Forecasting Macroeconomic Series h-steps Ahead,” Journal of
Econometrics 135, 499-526.
Mark, N.C. (1995), “Exchange Rates and Fundamentals: Evidence on Long-Horizon
Predictability,” American Economic Review 85(1), 201-218.
Mark, N.C. (2001), International Macroeconomics and Finance: Theory and Econometric
Methods, Malden, MA: Blackwell Publishers.
Mark, N.C. and D. Sul (2001), “Nominal Exchange Rates and Monetary Fundamentals:
Evidence from a Small Post-Bretton Woods Sample,” Journal of International Economics
54

53, 29-52.
Meese, R.A. and K.S. Rogo¤ (1983a), “Empirical Exchange Rate Models of the Seventies:
Do They Fit Out of Sample?,”Journal of International Economics 14(1-2), 3-24.
Meese, R.A. and K.S. Rogo¤ (1983b), “The Out-of-Sample Failure of Empirical Exchange
Rate Models: Sampling Error or Mis-speci…cation?,” in: Exchange Rates and International
Macroeconomics, Jacob Frenkel, eds., Chicago: NBER and University of Chicago Press.
Meese, R.A. and K.S. Rogo¤ (1988), “Was it Real? The Exchange Rate-Interest Di¤er-
ential Relation Over the Modern Floating Rate Period,”Journal of Finance 43, 923-948.
Meese, R.A. and A. Rose (1991), “An Empirical Assessment of Non-linearities in Models
of Exchange Rate Determination,”Review of Economic Studies 58, 603-619.
Melvin, M., J. Prins and D. Shand (2011), “Forecasting Exchange Rates: An Investor
Perspective,” in: Elliott, G. and A. Timmermann, Handbook of Economic Forecasting Vol.
2, Elsevier North-Holland.
Mincer, J. and V. Zarnowitz (1969), “The Evaluation of Economic Forecasts,” in J.
Mincer (ed.), Economic Forecasts and Expectations, New York: National Bureau of Economic
Research.
Mizrach, B.M. (1992), “Multivariate Nearest-Neighbor Forecasts of EMS Exchange Rates,”
Journal of Applied Econometrics 7, S151–S163.
Molodtsova, T. and D.H. Papell (2009), “Out-of-Sample Exchange Rate Predictability
with Taylor Rule Fundamentals,”Journal of International Economics, 77(2), 167-180.
Molodtsova, T. and D.H. Papell (2012), “Taylor Rule Exchange Rate Forecasting During
the Financial Crisis,”NBER International Seminar on Macroeconomics.
Molodtsova, T., A. Nikolsko-Rzhevskyy and D.H. Papell (2010), “Taylor Rules and the
Euro,”Journal of Money, Credit and Banking, 535-552.
Murray, C. and D.H. Papell (2002), “The Purchasing Power Parity Persistence Para-
digm,”Journal of International Economics 56, 1–19.
Mussa, M. (1976), “A Model of Exchange Rate Dynamics,” Scandinavian Journal of
Neely, C.J. (1997), “Technical Analysis in the Foreign Exchange Market: A Layman’s
Guide,”Federal Reserve Bank of St. Louis Review, September-October.
Neely, C.J. and S.R. Dey (2010), “A Survey of Announcement E¤ects on Foreign Ex-
change Returns,”Federal Reserve Bank of St. Louis Review 92(5), 417-463.
Neely, C.J. and L. Sarno (2002), “How Well Do Monetary Fundamentals Forecast Ex-
change Rates?”, Federal Reserve Bank of St. Louis Review, September-October.
55

Newey, W. and K.D. West (1987), “A Simple, Positive Semi-De…nite, Heteroskedasticity
and Autocorrelation Consistent Covariance Matrix,”Econometrica 55, 703–708.
Obstfeld, M. and K.S. Rogo¤ (1996), “Foundations of International Macroeconomics,”
Cambridge, MA: MIT Press.
Patton, A. and A. Timmermann (2006), “Testing Forecast Optimality Under Unknown
Loss,”Journal of American Statistical Association 102(480), 1172-1184.
Patton, A. and A. Timmermann (2007), “Properties of Optimal Forecasts under Asym-
metric Loss and Nonlinearity,”Journal of Econometrics 140, 884-918.
Pesaran, M.H. and A. Timmermann (1992), “A Simple Nonparametric Test of Predictive
Performance,”Journal of Business and Economic Statistics 10(4), 461-465.
Qi, M. and Y. Wu (2003), “Nonlinear Prediction of Exchange Rates with Monetary
Fundamentals,”Journal of Empirical Finance 10, 623-640.
Rapach, D.E. and M.E. Wohar (2002), “Testing the Monetary Model of Exchange Rate
Determination: New Evidence From a Century of Data,”Journal of International Economics
58, 359–385.
Rapach, D.E. and M.E. Wohar (2004), “Testing the Monetary Model of Exchange Rate
Determination: A Closer Look at Panels,” Journal of International Money and Finance
23(6), 867-895
Rapach, D.E. and M.E. Wohar (2006), “The Out-of-Sample Forecasting Performance of
Nonlinear Models of Real Exchange Rate Behavior,” International Journal of Forecasting
22(2), 341-361.
Rime D., L. Sarno and E. Sojli (2010), “Exchange Rate Forecasting, Order Flow and
Macroeconomic Information,”Journal of International Economics 80(1), 72-88.
Rogo¤, K. (1996), “The Purchasing Power Parity Puzzle,” Journal of Economic Litera-
ture XXXIV, 647-668.
Rogo¤, K. and V. Stavrakeva (2008), “The Continuing Puzzle of Short Horizon Exchange
Rate Forecasting,”NBER Working Paper 14071.
Rossi, B. (2005a), “Testing Long-Horizon Predictive Ability with High Persistence, and
the Meese-Rogo¤ Puzzle,”International Economic Review 46(1), 61-92.
Rossi, B. (2005b), “Optimal Tests for Nested Model Selection with Underlying Parameter
Instability,”Econometric Theory 21(5), 962-990.
Rossi, B. (2005c), “Con…dence Intervals for Half-Life Deviations from Purchasing Power
Parity,”Journal of Business and Economic Statistics 23(4), 432-442.
Rossi, B. (2006), “Are Exchange Rates Really Random Walks? Some Evidence Robust
56

to Parameter Instability,”Macroeconomic Dynamics 10(1), 20-38.
Rossi, B. (2007a), “Expectations Hypotheses Tests and Predictive Regressions at Long
Horizons,”Econometrics Journal 10(3), 1-26
Rossi, B. (2011), “Advances in Forecasting Under Instabilities,” in: Elliott, G. and A.
Timmermann, Handbook of Economic Forecasting Vol. 2, Elsevier-North Holland.
Rossi, B. and T. Sekhposyan (2011), “Understanding Models’Forecasting Performance,”
Journal of Econometrics 164(1), 158-172.
Rossi, B. and T. Sekhposyan (2012), “Conditional Predictive Density Evaluation in the
Presence of Instabilities”, mimeo.
Samuelson, P. (1964), “Theoretical Notes on Trade Problems,”Review of Economics and
Statistics 46, 145-154.
Sarantis, N. (1994), “The Monetary Exchange Rate Model in the Long Run: An Empir-
ical Investigation,”Weltwirtschaftliches Archiv 130, 698–711.
Satchell, S. and A. Timmermann (1995), “An Assessment of the Economic Value of
Non-linear Foreign Exchange Rate Forecasts”, Journal of Forecasting 14(6), 477-497.
Schinasi, G.J. and P.A.V.B. Swamy (1987), “The Out-of-Sample Forecasting Performance
of Exchange Rate Models when Coe¢ cients are Allowed to Change,”Journal of International
Money and Finance 8(3), 375-390.
Stock, J.H. and M.W. Watson (1993), “A Simple Estimator of Cointegrating Vectors in
Higher-Order Integrated Systems,” Econometrica 61(4), 783-820.
Stock, J.H. and M.W. Watson (1998), “Median Unbiased Estimation of Coe¢ cient Vari-
ance in a Time-Varying Parameter Model,” Journal of the American Statistical Association
93(441), 349-358.
Svensson, L.E.O. (2000), “Open Economy In‡ation Targeting,”Journal of International
Taylor, A.M. (2002), “A Century of Purchasing Power Parity,”Review of Economics and
Statistics 84(1), pages 139-150.
Taylor, J.B. (1993), “Discretion Versus Policy Rules in Practice,” Carnegie-Rochester
Conference Series on Public Policy 39, 195-214.
Taylor, M. P. and D.A. Peel (2000), “Nonlinear Adjustment, Long-Run Equilibrium and
Exchange Rate Fundamentals,”Journal of International Money and Finance 19(1), 33-53.
Taylor, M.P. and L. Sarno (1998), “The Behavior of Real Exchange Rates During the
Post-Bretton Woods Period”, Journal of International Economics 46(2), 281-312.
Terasvirta, T. (2006), “Forecasting Economic Variables with Nonlinear Models,” in G.
57

Elliott, C. Granger and A. Timmermann (eds.), Handbook of Economic Forecasting Vol. 1,
Elsevier-North Holland.
Theil, H. (1966), Applied Economic Forecasting, North Holland.
Timmermann, A. (2006), “Forecast Combinations,” in: Elliott, G., C. Granger and A.
Timmermann, Handbook of Economic Forecasting Vol. 1, Elsevier-North Holland North-
Holland.
Van Dijk, D. and P.H. Franses (2003), “Selecting a Nonlinear Time Series Model Using
Weighted Tests of Equal Forecast Accuracy,” Oxford Bulletin of Economics and Statistics
65, 727-744.
Wang, J. and J.J. Wu (2009), “The Taylor Rule and Interval Forecast for Exchange
Rates,”Federal Reserve Board International Finance Discussion Papers 963.
West, K.D. (1996), “Asymptotic Inference about Predictive Ability,” Econometrica 64,
1067-1084.
West, K.D. (2006), “Forecast Evaluation,”in: Elliott, G., C. Granger and A. Timmerman
(eds.), Handbook of Economic Forecasting Vol. 1, Elsevier-North Holland, 100-134.
West, K.D., H.J. Edison and D. Cho (1993), “A Utility-based Comparison of Some Models
of Exchange Rate Volatility,”Journal of International Economics 35, 23-46.
West, K.D., and M.W. McCracken (1998), “Regression-Based Tests of Predictive Ability,”
International Economic Review 39, 817-840.
Wieland, W. and M.H. Wolters (2011), “Forecasting and Policy Making,”in: Elliott, G.
and A. Timmermann (eds.), Handbook of Economic Forecasting, Volume 2, Elsevier North
Holland, forthcoming.
Wol¤, C.P (1987), “Time-Varying Parameters and the Out-of-Sample Forecasting Perfor-
mance of Structural Exchange Rate Models,” Journal of Business and Economic Statistics
5(1), 87-97.
Wright, J. (2008), “Bayesian Model Averaging and Exchange Rate Forecasting,”Journal
of Econometrics 146, 329-341.
58

Tables and Figures
Table 1. Literature Review: Predictors and Economic Models
Predictors (ft ) Economic Fundamentals Mnemonics
it i t Interest Rate Di¤erentials i
Ft st Forward Discount F
pt pt (Log) Price Di¤erentials p
t t In‡ation Di¤erentials
yt yt (Log) Output Di¤erentials y
mt mt (Log) Money Di¤erentials m
zt Productivity Di¤erentials z
bt bt Asset Di¤erentials b
ytgap ytgap Output Gap Di¤erentials y gap
nxat Net Foreign Assets nxa
CPt Commodity Prices CP
Model ft Mnemonics
UIRP (CIRP) it it ; (Ft st ) i; F
PPP pt pt or t t p;
Monetary Model with Flexible Prices (I) [(it it ) ; (yt yt ) ; (mt mt )]0 i; y; m
Monetary Model with Flexible Prices (II) [(yt yt ) ; (mt mt )]0 y; m
(or Frenkel-Bilson Model)
Monetary Model with Sticky Prices (I) [(it it ) ; (yt yt ) ; (mt mt ) ; (pt pt )]0 i; y; m; p
0
Monetary Model with Sticky Prices (II) [(it it ) ; (yt yt ) ; (mt mt ) ; ( t t )] i; y; m;
(or Dornbush-Frankel Model)
Model with Productivity Di¤erentials [(it it ) ; (yt yt ) ; (mt mt ) ; zt ]0 i; y; m; z
(or Balassa-Samuelson (1964) Model)
Portfolio Balance Model [(it it ) ; (bt bt )]0 i; b
(or Hooper and Morton (1982) Model)
gap 0
Taylor Rule Model [( t t ) ; (yt ytgap )] ; y gap
Net Foreign Asset Model nxat nxa
Commodity Prices CPt CP
59

Table 2. Literature Review: Models, Predictors and Data
A. Models and Predictors B. Data
Reference Linear Eq. Details Predictors Freq. Sample Countries
Single Equation Models
Abhyankar et al. (2005) L ECM calibr. coe¤. m,y M 1977:1-2000:12 CA,JP,UK
Alquist and Chinn (2008) L ECM estim. coe¤. i; (m,y , ,i);nxa Q 1970:1-2005:Q4 CA,UK,EU
Bacchetta et al. (2010) L contemp. real. fund. m,y ,u,i,CP M 1975:9-2008:9 UK,SWI,JP,GER,AU,CA,NZ
Berkowitz and Giorgianni (2001) L ECM calibr. coint. m,y Q 1973Q2-1991Q4 CA,GER,SWI,JP
Chen, Rogo¤ and Rossi (2010) L lag fund. CP Q varies by country AU,CAN,NZ,CHI,SA
Cheung et al. (2005) L di¤.; ECM estim. coint. i;p;(m,y ,i, );others(z ,b) Q 1973Q2-2000Q4 CAN,UK,GER,JP,SWI
Chinn (1991) NL levels,di¤,ECM non-parametric m,y ,i; ,wealth Q 1974Q1-1988Q4 GER,JP
Chinn and Meese (1995) L/NL ECM/LWR calibr. coe¤. m,y ,i; others( ,T B ,q ) M 1973:3-1990:11 GER,CA,UK,JP
Chinn and Moore (2012) L ECM estim. coint. order ‡ow M 1999:1-2001:1 EU,JP
Clark and West (2006) L lag fund. i M varies by country CAN,JP,SWI,UK
Della Corte et al. (2010) L lag fund. nxa Q (real t) 1973Q1-2007Q4 CAN,GER,UK,JP
Diebold and Nason (1990) NL univ. LWR s W 1973-1987 10 countries
Engel and Hamilton (1990) NL MS s,i Q 1973Q3-1988Q1 GER,FR,UK
Engel (1994) NL MS s Q 1973-1991 18 countries
Engel, Mark and West (2009) L ECM factor factor from s Q 1973-2007 17 countries
Faust et al. (2003) L ECM calibr. coint. y ,m; (m,y , ,i,b) Q (real t) 1973Q2-1991Q4 GER,JP,CAN,SWISS
Ferraro et al. (2011) L lag/contemp. CP D (real t) 1984:12-2010-11 CA,NOR,SA,AUS,CHI
Giacomini and Rossi (2010) L lag fund. ; y gap M 1973:1-2008:1 12 countries
Gourinchas and Rey (2007) L lag fund. nxa Q 1973:Q1-2004Q1 US e¤ective exch.
Groen (1999) L ECM m,y M 1973:4-1994:12 GER,FR,NET,CA
Inoue and Rossi (2012) L ECM ; y gap ; i M 1973:3-2008:1 10 countries
Kilian (1999) L ECM calibr. coint. m,y Q 1973Q2-1991Q4 CAN,GER,SWI,JP
Mark (1995) L ECM calibr. coint. m,y Q 1973Q2-1991Q4 CAN,GER,SWI,JP
Meese and Rogo¤ (1983a) L levels OLS,GLS,IV (m,y ,i);others( ,T B ,F ) M 1973:3-1981:6 UK,GER,JP,trade-w.
Meese and Rogo¤ (1983b) L contemp. calib. coe¤. (m,y ,i);others( ,T B) M 1973:3-1981:6 UK,GER,JP,trade-w.
Meese and Rose (1991) NL lag/contemp. LWR (m,y ,p);others(T B) M 1974-1987 CA,GER,JP,UK
Molodtsova et al. (2010) L lag fund. ; y gap Q (real t) 1991Q1-2007Q3 EU
Molodtsova and Papell (2009) L lag fund. ; y gap ;m,y ;i;p M varies by country 12 countries
60
Molodtsova and Papell (2012) L lag fund. ; y gap ,F CI Q (real t) 1992Q4-2012Q1 EU
Qi and Wu (2003) NL ECM neural net. m,y ,i M 1973:3-1997:7 JP,UK,CA,GER
Rapach and Wohar (2006) NL univ. ESTAR s M 1980:1-1994:12 UK,GER,FR,JP
Rime, Sarno and Sojli (2010) L ECM order ‡ows D 2004:2-2005:2 UK,EU,JP
Rogo¤ and Stavrakeva (2008) L ECM (m,y);( ; y gap );nxa Q 1973:1-2005:4 18 countries
Rossi (2005a) L ECM m,y Q 1973Q3-1998Q2 GER,JP
Rossi (2006) L/NL lag fund. TVP y , m,i M 1973:3-1998:12 CA,FR,GER,IT,JP
Rossi and Sekhposyan (2011) L lag fund. m, u, y , i, CP M 1975:9-2008:9 SWI,UK,CA,JP,GER
Schinasi and Swami (1987) L contemp. TVP-AR m,y ,i,T B M 1973:3-1980:3 UK,GER,JP
West, Edison and Cho (1993) L/NL univ. GARCH/NP s W 1973:3-1989:9 CA,FR,GER,JP,UK
Wol¤ (1987) L contemp. TVP-AR m,y; i, M 1973:3-1984:4 GER,JP,UK
Wu and Wang (2009) L lag fund. semi-param. ( ; y gap ; q);(m; y) M varies by country 12 countries
Multiple Eqs. Models
Canova (1993) L TVP-VAR i W 1979 to 1987 FR,SWI,GER,UK,JP
Clarida et al. (2003) NL VECM-MS F W 1979:1-1998:52 FR,GER,JP,UK
Clarida and Taylor (1997) L VECM F W 1977:1-1993:52 UK,GER,JP

Diebold et al. (1994) L VECM s D 3/1/1980-1/28/1985 CAN,FR,GER,IT,JP,SWI,UK
Diebold, Hahn and Tay (1999) L density s 30 min. 1996:1-1996:12 GER,JP
MacDonald and Taylor (1993) L VECM m; y ,i M 1976:1-1990:12 GER
Mizrach (1992) NL LWR s D 1986:1-1988:12 FR,IT,GER
Rapach and Wohar (2002) L VECM (m; y); i Y 115 years 14 countries
Wright (2008) L contemp/lag BMA p,y; i,z ,m,SP ,D,CA M,Q (real t) 1973-2005 CA,GER,JP,UK
Panel Models
Adrian, Etula and Shin (2011) L lag. fund. repo;comm.paper M 1993:1-2007:12 23 countries
Carriero et al. (2009) L large BVAR s M 1995:1-2008:4 33 countries
Cerra and Saxena (2010) L level,demean ECM; di¤. m; y Y (real t) 1960-2004 98 countries
Engel, Mark and West (2007) L VECM calibr. coe¤. (m; y; p); ( ; y gap ) Q 1973-2005 18 countries
Groen (2005) L ECM calibr. coint. m; y Q 1975Q1-2000Q4 CA,JP,GER
Mark and Sul (2001) L ECM m; y; p Q 1973Q1-1997Q1 19 countries
Rapach and Wohar (2004) L ECM m; y Q 1973Q1-1997Q1 18 countries
Table 2 (continued). Forecast Evaluation and Empirical Conclusions
C. Forecast Evaluation D. Successful Predictive Ability?
Reference Benchmark Estim. h Forec. Eval. Method
Single Equation Models
Abhyankar et al. (2005) VAR 3-48 mo. 1991:1 utility Yes
Alquist and Chinn (2008) RW roll 1,4,20 qrs. 1987Q2 CW No (except UIRP at long h and NFA for some countries at short h)
Bacchetta et al. (2010) RW roll 1 mo. varies MSFE No
Berkowitz and Giorgianni (2001) RW 1-16 qrs. 1981Q4 R2,DMW No
Chen, Rogo¤ and Rossi (2010) RW,AR roll 1 q. ENCNEW Yes
Cheung et al. (2005) RW roll(42, 59) 1,4,20 qrs 1987Q2 MSE,sign No
Chinn (1991) RW roll(48) 1, 4 qrs. 1986Q1 No
Chinn and Meese (1995) RW 1,12,36 mo. 1983:01 No (except monetary at long h for some countries)
Chinn and Moore (2012) RW 1-3-6 mo. 3 years CW Yes (but signi…cant only for one country)
Clark and West (2006) RW roll(120) 1 mo. varies DMW,CW Yes (for UIRP at short horizons)
Della Corte et al. (2010) roll(80) 1993Q1 utility-based Yes (for NXA)
Diebold and Nason (1990) RW 1-4 qrs. (iter.) 1986W1 No
Engel and Hamilton (1990) RW 1-4 qrs. (iter.) 1984Q1 MSE Yes
Engel (1994) RW,F 1-12 qrs. 1986 DMW Yes
Engel, Mark and West (2009) RW,Taylor 1-16 qrs. 1999 Yes, somewhat (at long h)
Faust et al. (2003) RW roll(40) 1 day No
Ferraro et al. (2011) RW ,RW W D roll 1 mo. varies DMW Yes
Giacomini and Rossi (2010) RW roll(50) 1-24 qrs. 1983:2 GR,CW Somewhat (it depends on time period)
Gourinchas and Rey (2007) RW 1-48 mo. 1978Q2 CW Yes
Groen (1999) RW rec 1 mo. 1981:10 DMW No
Inoue and Rossi (2012) RW roll 1-16 qrs. varies Yes
Kilian (1999) RW rec 1-16 qrs. 1981Q4 R2,DMW No
Mark (1995) RW rec 1,6,12 mo. 1981Q4 R2,DMW Yes (only at long h)
Meese and Rogo¤ (1983a) RW,VAR roll(93) 1,6,12 mo. 1976:11 MSFE,MAE No
Meese and Rogo¤ (1983b) RW,VAR 48 mo. 1976:11 MSFE,MAE No
Meese and Rose (1991) RWWD rec 1-4 qrs. 1984:1 MSFE Yes
Molodtsova et al. (2010) RW roll(26) 1999Q4 CW Yes
Molodtsova and Papell (2009) RW roll(120) 1 mo. 1982:3 CW Yes for Taylor; no for UIRP and monetary
61
Molodtsova and Papell (2012) RW roll(26) 1 q. 2007Q1 CW Yes
Qi and Wu (2003) roll(.75T) 1,6,12 mo. (iter) 1990:1 DMW No
Rapach and Wohar (2006) 1 to 3 mo. DMW,PIT No
Rime, Sarno and Sojli (2010) 1 day 2004:6 utility-based Yes
Rogo¤ and Stavrakeva (2008) varies various Mostly no
Rossi (2005a) 1-12qrs. DMW Yes, based on GC-robust tests; somewhat, based on TVP models
Rossi (2006) RW roll,rec 1 mo. D M W ,E N C N E W Yes (for some countries)
Rossi and Sekhposyan (2011) RW roll 1-12mo. varies DMW No
Schinasi and Swami (1987) 15 mo. 1980:4 MSFE,MAE Yes (but no statistical signi…cance analysis)
West, Edison and Cho (1993) utility-based Yes
Wol¤ (1987) RW rec 1-24 mo. (iter) 1981:7 MSFE,MAE Somewhat (only for one country)
Wu and Wang (2009) RW roll(200) 1-12 mo. (iter) interval,GW Yes
Multiple Eqs. Models
Canova (1993) RW 1-52 w. (iter) 1986 Theil Yes
Clarida et al. (2003) RW 4-52 w. 1996:1 DMW Yes
Clarida and Taylor (1997) RW rec 1-12 mo. 1977:1 MSFE,MAE Yes

Diebold et al. (1994) 1-26 days No
Diebold, Hahn and Tay (1999) absolute PIT n.a.
MacDonald and Taylor (1993) RW rec 1-12 mo. (iter) 1989-1990 MSFE Yes
Mizrach (1992) RW 5 days 1974:1 MSFE,MAE No
Rapach and Wohar (2002) RW rec ENCNEW Somewhat (depends on country)
Wright (2008) RW 3-12 mo. 1990 MSFE Yes (although magnitude of success is marginal)
Panel Models
Adrian, Etula and Shin (2011) RW,AR(1) roll(48) 1-12 mo. 1997:1 CW Yes for advanced countries, somewhat for emerging countries
Carriero et al. (2009) RW roll(84) 1-12 mo. 2001:1 GW Yes
Cerra and Saxena (2010) RW ,RW W D rec 1-5 yrs. 1984 various Yes
Engel, Mark and West (2007) RW rec 1-16 qrs. 1983 CW Yes (but not always)
Groen (2005) RW rec 1-16 (iter) 1989Q1 bootstrap Yes (at long horizons)
Mark and Sul (2001) RW rec 1, 16 qrs. 1983Q2 MSFE,Theil Yes
Rapach and Wohar (2004) heter. rec(50) 1-16 qrs. n.a.
Table 3. Available Sample Sizes, By Country and Model
Models in Monthly Data
Exchange Rate Di¤erential Interest Rate Di¤erential CPI Di¤erential Money Di¤erential Output Di¤erential
Country Start End N.Obs. Start End N.Obs. Start End N.Obs. Start End N.Obs. Start End N.Obs.
Australia 1957:02 2011:06 653 1969:08 2011:06 503 n.a. n.a. n.a. 1975:03 2011:06 436 1957:02 2007:09 608
Austria 1957:02 1998:12 503 1967:02 1998:12 383 1995:02 2011:06 197 1997:10 2011:06 165 1957:02 2011:05 652
Belgium 1957:02 1998:12 503 1957:02 1999:01 504 1957:02 2011:06 653 2002:02 2011:06 113 1957:02 2011:03 650
Canada 1957:02 2011:06 653 1975:02 2011:06 437 1957:02 2011:06 653 1975:04 2011:06 435 1957:02 2011:05 652
Denmark 1957:02 2011:06 653 1972:02 2011:06 469 1980:02 2011:06 377 1991:02 2011:06 245 1957:02 2011:05 652
Finland 1957:02 1998:12 503 1978:01 2011:06 402 1960:02 2011:06 617 1960:02 2011:06 617 1957:02 2011:06 653
France 1957:02 1998:12 503 1964:02 1999:03 418 1990:02 2011:06 257 1978:01 2011:06 402 1957:02 2011:05 652
Germany 1957:02 1998:12 503 1960:02 2011:06 617 1957:02 2011:06 653 1973:02 2011:06 461 1958:02 2011:05 640
Greece 1957:02 2000:12 527 1998:02 1999:10 021 1959:02 2011:06 629 1980:02 2011:06 377 1957:02 2011:05 652
Ireland 1957:02 1998:12 503 1972:04 2011:06 450 1969:02 2011:06 509 1999:02 2011:06 149 1976:02 2011:05 424
Italy 1957:02 1998:12 503 1971:02 2011:06 485 1957:02 2011:06 653 1980:02 2011:06 377 1957:02 2011:05 652
Japan 1957:02 2011:06 653 1957:02 2011:06 653 2000:02 2011:06 137 1960:02 2011:06 617 1957:02 2011:06 653
New Zealand 1957:02 2011:06 653 1985:02 2011:06 317 n.a. n.a. n.a. 1977:04 2011:06 411 1987:02 2007:12 251
Portugal 1957:02 1998:12 503 1983:02 2000:03 206 1960:02 2011:06 617 1997:10 2011:06 165 1957:02 2006:04 591
Spain 1957:02 1998:12 503 1974:02 2011:06 449 1957:02 2011:06 653 1997:10 2011:06 165 1961:02 2011:05 604
Sweden 1957:02 2011:06 653 1966:01 2011:06 546 1960:02 2011:06 617 1998:02 2011:06 161 1957:02 2011:05 652
Switzerland 1957:02 2011:06 653 1975:10 2011:06 429 1957:02 2011:06 653 1985:01 2011:06 318 1995:02 2007:12 155
UK 1957:02 2011:06 653 1972:02 2011:06 473 1988:02 2011:06 281 1986:10 2011:06 297 1957:02 2011:05 652
62
Models in Quarterly Data
Int. Rate Di¤. CPI Di¤. TB Di¤. Output Di¤. CA Di¤.
Country Start End N.Obs. Start End N.Obs. Start End N.Obs. Start End N.Obs. Start End N.Obs.
Australia 1969:04 2011:02 167 1957:02 2011:02 217 1973:02 2010:04 151 1959:04 2011:02 207 1973:02 2010:04 151
Austria 1967:02 1998:04 127 1957:02 2011:02 217 1973:02 2011:01 152 1988:02 2011:01 092 1973:02 2011:01 152
Belgium 1957:02 1998:04 167 1957:02 2011:02 217 2002:02 2011:01 036 1980:02 2011:02 125 2002:02 2011:01 036
Canada 1975:02 2011:02 145 1957:02 2011:02 217 1973:02 2011:01 152 1961:02 2011:02 201 1973:02 2011:01 152
Denmark 1972:02 2011:02 155 1980:02 2011:02 125 1975:02 2011:01 144 1991:02 2011:02 081 1975:02 2011:01 144
Finland 1978:02 2011:02 133 1960:02 2011:02 205 1975:02 2011:01 144 1975:02 2011:02 145 1975:02 2011:01 144
France 1957:02 1999:01 168 1990:02 2011:02 085 1975:02 2010:04 143 1957:02 2011:02 217 1975:02 2010:04 143
Germany 1957:02 2011:02 217 1957:02 2011:02 217 1973:02 2011:01 152 1991:02 2011:02 081 1973:02 2011:01 152
Greece 1998:02 1999:03 006 1959:02 2011:02 209 1976:02 2011:01 135 2000:02 2011:01 044 1976:02 2011:01 135
Ireland 1973:02 2011:02 147 1969:02 2011:02 169 1981:02 2011:01 120 1997:02 2011:01 056 1981:02 2011:01 120
Italy 1971:02 2011:02 161 1957:02 2011:02 217 1973:02 2011:01 152 1981:02 2011:02 121 1973:02 2011:01 152
Japan 1957:02 2011:02 217 2000:02 2011:02 045 1977:02 2011:01 136 1980:02 2011:02 125 1977:02 2011:01 136
New Zealand 1985:02 2011:02 105 1957:02 2011:02 217 1980:02 2011:01 124 1987:03 2011:01 095 1980:02 2011:01 124
Portugal 1981:02 2000:01 076 1960:02 2011:02 205 1975:02 2011:01 144 1995:02 2011:02 065 1975:02 2011:01 144

Spain 1974:02 2011:02 149 1957:02 2011:02 217 1975:02 2011:01 144 1995:02 2011:02 065 1975:02 2011:01 144
Sweden 1966:02 2011:02 181 1960:02 2011:02 205 1975:02 2011:01 144 1993:02 2011:02 073 1975:02 2011:01 144
Switzerland 1976:01 2011:02 142 1957:02 2011:02 217 1999:02 2011:01 048 1980:02 2011:02 125 1999:02 2011:01 048
UK 1972:02 2011:02 157 1988:02 2011:02 093 1973:02 2011:01 152 1957:02 2011:02 217 1973:02 2011:01 152
Table 4. Relative Models’Forecasting Ability (Single Eq. Models)
A. UIRP Model B. PPP Model C. Monetary Model D. Monet. ECM Model E. Taylor Model F. Portfolio Balance
Country GC RMSFE DMW CW GC RMSFE DMW CW GC RMSFE DMW CW GC RMSFE DMW CW GC RMSFE DMW CW GC RMSFE DMW CW
h=1 Month h=1 Month h=1 Month h=1 Month h=1 Month h=1 Quarter
Australia 0.44 1.00 0.52 0.58 n.a. n.a. n.a. n.a. 0.06 1.00 0.51 0.27 0.00 1.01 0.55 0.89 n.a. n.a. n.a. n.a. 0.66 1.06 0.56 0.31
Austria 0.10 1.01 0.53 0.48 0.27 1.07 0.74 0.99 n.a. n.a. n.a. n.a. n.a. n.a. n.a. n.a. 0.00 0.98 0.54 0.14 1.00 1.06 0.61 0.77
Belgium 0.26 1.01 0.54 0.56 1.00 1.01 0.56 0.91 n.a. n.a. n.a. n.a. n.a. n.a. n.a. n.a. 0.82 1.00 0.48 0.33 n.a. n.a. n.a. n.a.
Canada 0.00 1.00 0.51 0.40 0.09 0.99 0.45 0.00 1.00 1.01 0.54 0.88 0.01 1.00 0.57 0.99 0.22 0.99 0.46 0.00 0.30 1.03 0.55 0.65
Denmark 1.00 1.01 0.56 0.95 0.11 1.00 0.52 0.55 0.34 1.01 0.56 0.58 0.05 1.00 0.52 0.50 0.35 1.00 0.52 0.28 0.00 1.04 0.59 0.72
Finland 0.63 1.02 0.57 0.86 0.50 1.00 0.53 0.76 0.38 1.01 0.55 0.72 0.39 1.00 0.51 0.35 0.48 1.00 0.48 0.15 0.00 1.08 0.64 0.84
France 0.78 1.02 0.54 0.75 0.42 1.02 0.59 0.88 0.72 1.01 0.55 0.73 0.02 1.01 0.54 0.50 0.66 1.02 0.54 0.60 0.00 1.27 0.70 0.77
Germany 0.00 1.01 0.53 0.47 0.82 1.00 0.52 0.57 0.37 1.01 0.53 0.34 0.00 1.00 0.54 0.76 0.51 0.99 0.45 0.03 0.01 1.09 0.62 0.85
Greece 0.00 1.00 0.51 n.a. 0.13 1.00 0.49 0.02 0.54 0.99 0.48 0.00 0.14 1.00 0.52 0.53 0.00 0.99 0.47 0.00 n.a. n.a. n.a. n.a.
Ireland 0.00 1.03 0.54 0.76 0.00 1.01 0.54 0.58 n.a. n.a. n.a. n.a. n.a. n.a. n.a. n.a. 0.06 1.01 0.54 0.44 0.00 1.12 0.58 0.45
Italy 0.28 1.02 0.55 0.80 0.00 1.01 0.55 0.76 0.08 1.00 0.51 0.28 0.00 1.01 0.56 0.63 0.03 1.00 0.47 0.04 0.08 1.07 0.58 0.77
Japan 0.05 1.01 0.54 0.81 0.47 1.01 0.51 0.31 0.26 1.01 0.52 0.25 0.06 1.00 0.52 0.59 0.48 0.96 0.42 0.03 0.00 1.24 0.65 0.31
New Zealand 1.00 1.01 0.53 0.53 n.a. n.a. n.a. n.a. 0.01 1.00 0.50 0.32 0.00 1.01 0.57 0.94 n.a. n.a. n.a. n.a. 0.02 1.04 0.56 0.50
Portugal 0.00 1.02 0.58 0.91 0.00 1.01 0.56 0.77 n.a. n.a. n.a. n.a. n.a. n.a. n.a. n.a. 0.01 1.01 0.53 0.25 0.00 1.15 0.63 0.76
Spain 0.74 1.02 0.56 0.81 0.07 1.01 0.58 0.97 n.a. n.a. n.a. n.a. n.a. n.a. n.a. n.a. 0.02 1.00 0.49 0.08 0.00 1.04 0.55 0.44
Sweden 0.00 1.04 0.53 0.88 0.19 1.00 0.50 0.30 0.17 1.04 0.62 0.98 0.00 0.99 0.46 0.13 0.24 0.99 0.48 0.01 0.00 1.03 0.55 0.38
Switzerland 0.45 1.01 0.55 0.84 0.36 1.00 0.50 0.38 0.78 1.01 0.58 0.91 0.02 1.02 0.63 0.99 0.44 1.01 0.55 0.25 0.00 1.12 0.57 0.09
63
UK 1.00 1.01 0.58 1.00 1.00 1.01 0.54 0.70 0.61 1.01 0.58 0.89 0.20 1.01 0.55 0.61 1.00 1.00 0.49 0.10 0.20 1.03 0.57 0.29
h=4 Years h=4 Years h=4 Years h=4 Years h=4 Years h=4 Years
Australia n.a. n.a. n.a. n.a. n.a. n.a. n.a. n.a. 0.21 1.00 0.52 0.29 0.57 1.01 0.57 0.93 n.a. n.a. n.a. n.a. 0.05 1.03 0.60 0.86
Austria 0.53 1.00 0.53 0.67 n.a. n.a. n.a. n.a. n.a. n.a. n.a. n.a. n.a. n.a. n.a. n.a. n.a. n.a. n.a. n.a. 0.00 1.05 0.59 0.11
Belgium 1.00 1.01 0.53 0.50 0.47 1.01 0.58 0.97 n.a. n.a. n.a. n.a. n.a. n.a. n.a. n.a. 0.89 1.00 0.51 0.49 n.a. n.a. n.a. n.a.
Canada 0.43 1.01 0.53 0.36 0.28 1.00 0.52 0.40 1.00 1.00 0.53 0.73 0.50 1.00 0.51 0.38 0.12 0.99 0.48 0.00 0.00 1.03 0.57 0.65
Denmark 0.71 1.00 0.52 0.43 0.72 1.00 0.51 0.44 1.00 1.02 0.57 0.70 0.29 1.02 0.57 0.71 1.00 1.01 0.49 0.36 0.26 1.04 0.59 0.59
Finland 0.31 1.01 0.55 0.78 0.83 1.00 0.55 0.84 0.55 1.01 0.55 0.50 0.79 1.01 0.54 0.66 0.10 1.02 0.51 0.46 0.00 1.09 0.65 0.77
France 1.00 1.02 0.57 0.76 0.53 1.04 0.61 0.78 0.21 1.02 0.55 0.73 0.51 1.01 0.55 0.80 0.23 n.a. n.a. n.a. 0.00 1.21 0.65 n.a.
Germany 0.82 1.01 0.54 0.60 0.04 1.01 0.54 0.69 0.22 1.01 0.54 0.53 0.48 1.01 0.56 0.86 0.40 1.01 0.49 0.28 0.01 1.11 0.61 0.60
Greece 0.77 1.01 0.53 0.49 0.22 1.00 0.49 0.03 0.43 1.00 0.49 0.04 0.78 1.00 0.51 0.16 0.21 1.00 0.50 0.01 n.a. n.a. n.a. n.a.
Ireland 1.00 1.02 0.56 0.85 0.40 1.01 0.56 0.83 n.a. n.a. n.a. n.a. n.a. n.a. n.a. n.a. 0.04 1.03 0.57 0.45 0.00 1.11 0.71 0.52
Italy 1.00 1.02 0.56 0.70 0.04 1.00 0.51 0.36 0.35 1.04 0.62 0.97 0.31 1.03 0.59 0.80 0.02 1.01 0.50 0.33 0.00 1.06 0.56 0.42
Japan 0.64 1.01 0.55 0.85 0.22 1.02 0.57 0.64 0.13 1.01 0.52 0.36 0.39 1.01 0.53 0.80 0.00 n.a. n.a. n.a. 0.00 1.09 0.61 0.22
New Zealand 0.64 1.01 0.55 0.71 n.a. n.a. n.a. n.a. 0.22 1.01 0.54 0.72 0.89 1.00 0.52 0.28 n.a. n.a. n.a. n.a. 0.02 1.06 0.63 0.90
Portugal 1.00 1.02 0.58 0.84 0.00 1.01 0.53 0.47 n.a. n.a. n.a. n.a. n.a. n.a. n.a. n.a. 0.00 1.12 0.55 0.20 0.00 1.06 0.59 0.28
Spain 0.70 1.02 0.55 0.61 0.05 1.00 0.53 0.63 n.a. n.a. n.a. n.a. n.a. n.a. n.a. n.a. 0.00 1.00 0.45 0.05 0.00 1.06 0.58 0.41

Sweden 0.44 1.04 0.54 0.75 0.59 1.00 0.53 0.46 0.80 1.02 0.59 0.62 1.00 1.02 0.59 0.81 0.21 0.99 0.50 0.02 0.02 1.04 0.59 0.69
Switzerland 0.19 1.01 0.53 0.46 0.13 1.00 0.52 0.40 0.39 1.01 0.54 0.54 1.00 1.00 0.52 0.27 0.49 n.a. 0.63 0.17 0.00 1.04 0.60 0.60
UK 0.24 1.01 0.57 0.92 0.63 1.00 0.50 0.21 0.56 1.00 0.54 0.67 1.00 1.01 0.54 0.68 1.00 1.00 0.47 0.11 0.61 1.07 0.63 0.91
Table 5. Relative Models’Forecasting Ability (Multiple Eqs. Models)
A. Monetary Panel Model B. BMA Model
Country RMSE DMW CW RMSE DMW CW
h=1 Month h=1 Month
h=1 Month
Australia 1.01 0.53 0.34 1.00 0.53 0.72
Austria 1.30 0.90 n.a. n.a. n.a. n.a.
Belgium n.a. n.a. n.a. n.a. n.a. n.a.
Canada 1.01 0.53 0.63 1.00 0.53 0.83
Denmark 1.02 0.61 0.99 1.00 0.53 0.66
Finland 1.01 0.55 0.78 1.01 0.57 0.91
France 1.00 0.52 0.31 1.00 0.51 0.38
Germany 1.01 0.53 0.46 1.00 0.51 0.55
Greece 1.00 0.50 0.02 1.09 0.67 n.a.
Ireland n.a. n.a. n.a. n.a. n.a. n.a.
Italy 1.00 0.51 0.29 1.00 0.53 0.59
Japan 1.01 0.53 0.62 1.00 0.49 0.21
New Zealand 1.01 0.54 0.44 1.00 0.54 0.78
Portugal 1.13 0.71 n.a. n.a. n.a. n.a.
Spain 1.01 0.56 n.a. n.a. n.a. n.a.
Sweden 1.00 0.54 0.69 1.01 0.57 0.89
Switzerland 1.03 0.61 0.97 1.00 0.52 0.56
64
U.K. 1.01 0.55 0.75 1.00 0.52 0.53
h=4 Years
Australia 1.35 0.70 0.31 1.00 0.48 0.18
Austria n.a. n.a. n.a. n.a. n.a. n.a.
Belgium n.a. n.a. n.a. n.a. n.a. n.a.
Canada 1.34 0.70 1.00 1.00 0.53 0.84
Denmark 1.49 0.76 1.00 1.01 0.55 0.83
Finland 1.84 0.80 0.00 1.01 0.55 0.87
France 1.72 0.78 1.00 1.00 0.54 0.79
Germany 1.75 0.76 1.00 1.00 0.52 0.50
Greece 1.14 0.60 0.00 n.a. n.a. n.a.
Ireland n.a. n.a. n.a. n.a. n.a. n.a.
Italy 1.11 0.55 0.00 1.00 0.54 0.73
Japan 1.79 0.74 0.50 1.00 0.52 0.59
New Zealand 1.52 0.69 1.00 1.00 0.49 0.27
Portugal n.a. n.a. n.a. n.a. n.a. n.a.
Spain n.a. n.a. n.a. n.a. n.a. n.a.

Sweden 1.10 0.61 0.02 1.00 0.49 0.21
Switzerland 1.27 0.77 0.87 1.00 0.49 0.20
U.K. 1.60 0.82 1.00 1.00 0.48 0.16
Notes to the Tables.
Note to Table 1. The table reports the name of the model "Model", the fundamental
predictors used in the model ("ft ") and the mnemonics used to refer to these fundamentals
in Table 2.
Note to Table 2, Panel A. "L" denotes linear model and "NL" denotes non-linear model.
Mnemonics for the predictors are as in Table 1. If a paper considers several models at
the same time, their predictors are separated by a semicolon; for example, a paper that
considers both UIRP and the monetary model with interest rates will be characterized by:
i; (m; ; y; i). In addition, "SP " denotes stock prices; "q" denotes the real exchange rate;
"CP " denotes commodity prices (either oil price or the commodity price); "D" denotes the
dividend yield; "CA" denotes the current account; "T B" denotes the trade balance; "s" is
the lagged exchange rate; "F CI" denotes Financial Condition Indices; "r" denotes the real
interest rate; "u" denotes unemployment; a " " before a variable denotes the …rst di¤erence
of that variable. Note that Cheung et al. (2005) also consider the model with productivity
(m; y; i; z) as well as a model with net foreign assets (p; r; b; terms of trade, net foreign assets
and the relative price of non-tradeables). "Calibr. coe¤." means calibrated coe¢ cients;
"estim. coe¤." denotes estimated coe¢ cients; "calibr. coint." means calibrated values for
the cointegration vector; "estim. coint." denotes estimated cointegration vector; "di¤."
means that the model is estimated with variables in di¤erences; "semi-param." means that
the model is estimated with semi-parametric methods; "contemp." denotes contemporaneous
predictors; "lag fund." denotes lagged fundamentals; "real. fund." means that the model uses
ex-post, realized values of the contemporaneous fundamentals for prediction; "univ." denotes
a univariate model (with possible lags of exchange rates and no fundamentals); "ECM"
denotes Error Correction Model; "NP" denotes non-parametric; "LWR" denotes locally
weighted regression; "MS" denotes Markov Switching model; "neural net." denotes neural
network model. Canova’s (1993) model includes stochastic volatility; in Cheung, Chinn and
Pascual (2005), the contemporaneous fundamental model is estimated with OLS whereas
the cointegrating vector in the ECM model is estimated via the Johansen procedure; Wright
(2008) also includes the ratio of the current account to output among the regressors; in Adrian
et al. (2011), the repo- and commercial paper variables (here labeled "repo;comm.paper")
are detrended over the full sample.
Note to Table 2, Panel B. "Freq." denotes the frequency of the data: "Y" for yearly,
65

"Q" for quarterly, "M" for monthly, "D" for daily, "W" for weekly, and "30 min." for 30-
minutes data. "real t" indicates that the study uses real-time vintages of data. "varies by
country" means that typically the sample is 1973:3-2006:6, but for EU countries it may stop
in 1998; "trade-w." denotes the trade-weighted exchange rate; "US e¤ective exch." denotes
the US e¤ective exchange rate. Country mnemonics are as follows: "JP" is Japan, "GER" is
Germany, "CA" is Canada, "UK" is the United Kingdom, "EU" is the Euro area, "SWI" is
Switzerland, "AU" is Australia, "NZ" is New Zealand, "CHI" is Chile, "SA" is South Africa,
"FR" is France, "NET" is the Netherlands, "IT" is Italy; "NOR" is Norway. Molodtsova
et al. (2010) use quarterly data interpolated from annual data; Carriero et al. (2009) use
average exchange rates rather than end-of-period data.
Note to Table 2, Panel C. The benchmark models include "Taylor" (the Taylor-rule
model), the random walk ("RW"), the random walk with drift ("RWWD"), the forward
discount ("F"), the autoregressive ("AR"), the vector autoregressive ("VAR") and the het-
erogeneous coe¢ cient panel model ("heter."). "Absolute" denotes cases in which there is
no benchmark model and the forecast evaluation is based on absolute (rather than relative)
terms. "Estim" denotes the parameter estimation method, and includes rolling ("roll") and
recursive ("rec") estimation; when available, the size of the rolling estimation window is
included in parentheses; "T " denotes the total sample size. "h" is the forecast horizon;
"1-16 qrs." typically denotes 1,4,8,12,16 quarters horizons; "(iter.)" denotes that forecasts
at horizons bigger than one are made with an iterated method; otherwise it is assumed
that the forecast method is direct. "Forec." denotes the starting date of the sample used
for out-of-sample forecast evaluation (the end of the forecast sample is typically the end
of the sample from Table 3); "varies" means that various window sizes have been used so
several out-of-sample periods have been investigated; Alquist and Chinn (2008) also consider
1999Q1-20005:Q4; Chinn and Meese (1995) also consider 1987-1990:11; Groen (1999) also
considers 1981:10-1991:12; Meese and Rogo¤ (1983a) also consider 1978:11-1981:6; Mizrach
(1992) also considers 1974:1-1979:3; Cheung, Chinn and Pascual (2005) also consider post
1982; sample information for daily data is only approximate. "Eval. Method" denotes the
method used for the forecast evaluation; "DMW" denotes the Diebold and Mariano (1995)
and West (1996) test; "CW" denotes the Clark and West (2006) test; "ENCNEW" denotes
the Clark and McCracken (2001) test; "GR" denotes the Giacomini and Rossi (2010) test;
"Theil" denotes Theil’s (1966) U statistic"; "GW" denotes the Giacomini and White (2006)
test; "sign" denotes the direction of change statistics; "utility-based" denotes utility-based
statistics; "MSFE" denotes the Mean Square Forecast Error; "MAE" denotes the Mean
66

Absolute Error; "R2" denotes the R-square statistic; "interval" denotes methods based on
interval evaluation; "PIT" denotes the Probability Integral Transform; "various" indicates
that several forecast evaluation methods have been used. Note that Rogo¤ and Stavrakeva
(2008) consider all of the following statistics: CW, Theil, DMW, ENCNEW; Cerra and
Saxena (2010) consider Theil, sign, DMW, CW.
Note to Table 2, Panel D. The information in the panel succintly summarizes whether
the author(s)’preferred speci…cation beats the benchmark and provides evidence in favor of
successful predictive ability.
Note to Table 3. The tables report the sample sizes available for each predictor and each
country. "Start" denotes the start of the sample; "End" denotes the end of the sample; "N.
Obs." denotes the total number of observations.
Note to Tables 4 and 5. The tables report p-values of the following tests: Rossi’s (2005)
Granger-causality robust ("GC"), Diebold and Mariano (1995) and West (1996) ("DMW")
and Clark and West (2006) ("CW"); the benchmark model in the latter two tests is the
random walk without drift. "h" denotes the forecast horizon. "RMSFERW " denotes the
root mean squared error of the random walk model; "RMSFER" denotes the ratio of the
root mean squared forecast error of the model relative to that of the random walk without
drift: values smaller than unity denote that the model forecasts better than the random
walk benchmark.
Notes to the Figures.

Notes to Figures 1, 2, 5 and 6. The …gures plot Giacomini and Rossi’s (2010) Fluctuation
test statistic (solid line) along with its critical value (dotted line) for selected countries,
models and forecast horizons. The x-axis denotes time. When indicated, R denotes the
window size used for parameter estimation; if not indicated, R equals half of the total sample
size. The window size used to smooth out-of-sample predictive ability in the Fluctuation
test is a third of the total number of forecasts.
Notes to Figure 3 and 4. The …gures plot Inoue and Rossi’s (2012) sup-test statistic
(solid line) along with its critical value (dotted line) for selected countries, models and
forecast horizons. The x-axis denotes the window size, R:
Notes to Figure 7. The …gure is a scatterplot of the p-values of Clark and West’s (2006)
test at short-horizons (on the x-axis) and at long-horizons (on the y-axis) for several predic-
tors, denoted by di¤erent markers, across multiple countries (each point denotes a country).
P-values smaller than 0.05 denote statistical signi…cance at the 5% level.
67

Figure 1. Fluctuation Test (Monetary-ECM Model, Short-Horizon)
MonECM for Canada at h=1 months MonECM for France at h=1 months
4 5
-2 0
Fluctuation test
Fluctuation test
-4
-6
-8 -5
-10
-12
-14 -10
1998 2000 2002 2004 2006 2008 2010 2012 1991 1992 1993 1994 1995 1996 1997 1998 1999
Time Time
MonECM for Switzerland at h=1 months MonECM for Germany at h=1 months
4 4
2
2
0
0
-2
Fluctuation test
Fluctuation test
-2
-4
-4
-6
-6
-8
-8
-10
-10 -12
2003.5 2004 2004.5 2005 2005.5 2006 2006.5 2007 2007.5 2008 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999
Time Time
MonECM for Japan at h=1 months MonECM for U.K. at h=1 months
5 4
0 0
-2
Fluctuation test
Fluctuation test
-5 -4
-6
-10 -8
-10
-15 -12
1998 2000 2002 2004 2006 2008 2010 2012 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012
Time Time
68

Figure 2. Fluctuation Test (Monetary-ECM Model, Long-Horizon)
10 4
8
2
6
0
4
-2
Fluctuation test
Fluctuation test
2
0 -4
-2
-6
-4
-8
-6
-10
-8
-10 -12
2000 2002 2004 2006 2008 2010 2012 1993 1994 1995 1996 1997 1998 1999
Time Time
4 5
2 0
0 -5
Fluctuation test
Fluctuation test
-2 -10
-4 -15
-6 -20
-8 -25
2007 2007.5 2008 2008.5 2009 2009.5 2010 2010.5 2011 2011.5 1991 1992 1993 1994 1995 1996 1997 1998 1999
Time Time
MonECM for U.K. at h=48 months

5
MonECM for Japan at h=48 months
4
-2 0
Fluctuation test
Fluctuation test
-4
-6
-8
-5
-10
-12
-14
-16 -10
1998 2000 2002 2004 2006 2008 2010 2012 2004 2005 2006 2007 2008 2009 2010 2011 2012
Time Time
69

Figure 3. Robustness to Window Size (Monetary-ECM Model, Short-Horizon)
6 15
4
10
2
5
Inoue-Rossi test
Inoue-Rossi test
0
-2
-5
-4
-10
-6
-8 -15
50 100 150 200 250 300 350 400 20 40 60 80 100 120 140 160 180 200 220
R R
10 5
5 0
0 -5
Inoue-Rossi test
Inoue-Rossi test
-5 -10
-10 -15
-15 -20
20 40 60 80 100 120 140 0 50 100 150 200 250 300
R R
16 10
14
5
12
0
10
Inoue-Rossi test
8
Inoue-Rossi test
-5
-10
4
2 -15
-20
-2
-4 -25
50 100 150 200 250 300 350 400 0 50 100 150 200 250
R R
70

Figure 4. Robustness to Window Size (Monetary-ECM Model, Long-Horizon)
10 5
5 0
0 -5
Inoue-Rossi test
Inoue-Rossi test
-5 -10
-10 -15
-15 -20
-20 -25
50 100 150 200 250 300 350 20 40 60 80 100 120 140 160 180
R R
5 5
0
0
-5
-5
Inoue-Rossi test
Inoue-Rossi test -10
-10
-15
-15
-20
-20
-25
-25 -30
20 30 40 50 60 70 80 90 100 110 120 20 40 60 80 100 120 140 160 180 200 220
R R
5 5
-5
0
-10
Inoue-Rossi test
Inoue-Rossi test
-15
-20
-5
-25
-30
-35 -10
50 100 150 200 250 300 350 20 40 60 80 100 120 140 160 180 200 220
R R
71

Figure 5. Interaction between Window Size and Out-of-Sample Period –Germany
MonECM for Germany at h=48 months MonECM for Germany at h=48 months
10 5
5
0
Fluctuation test
Fluctuation test
0
-5
-5
-10
-10
-15 -15
1985 1990 1995 2000 1985 1990 1995 2000
Time (R=50) Time (R=70)
MonECM for Germany at h=48 months MonECM for Germany at h=48 months
5 5
0 0
Fluctuation test
Fluctuation test
-5 -5
-10 -10
-15 -15
-20 -20
1985 1990 1995 2000 1990 1992 1994 1996 1998 2000
72

Figure 6. Interaction between Window Size and Out-of-Sample Period –Japan
MonECM for Japan at h=48 months MonECM for Japan at h=48 months
5 5
0 0
-5
Fluctuation test
Fluctuation test
-5
-10
-10
-15
-15
-20
-25 -20
-30 -25
1990 1995 2000 2005 2010 2015 1990 1995 2000 2005 2010 2015
MonECM for Japan at h=48 months MonECM for Japan at h=48 months
5 5
0 0
Fluctuation test
Fluctuation test
-5 -5
-10 -10
-15 -15
-20 -20
1990 1995 2000 2005 2010 2015 1990 1995 2000 2005 2010 2015
73

Figure 7. Out-of-sample Predictive Ability of Economic Models
1
UIRP
0.9 PPP
Monetary
0.8 Taylor-rule
Monetary Panel
P-value of CW Test at h=48
0.7 Monetary ECM
0.6
0.5
0.4
0.3
0.2
0.1
0
0 0.2 0.4 0.6 0.8 1
P-value of C W Test at h=1
Figure 8. In-sample Predictive Ability of Economic Models

1
UIRP
PPP
0.9 Monetary
Taylor-rule
Monetary ECM
0.8
P-value of GC-robust Test at h=48
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
P-value of GC-robust Test at h=1
74

Exchange Rate Predictability Rossi

Uploaded by

Copyright:

Available Formats

Exchange Rate Predictability Rossi

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Exchange Rate Predictability Rossi

Uploaded by

Copyright:

Available Formats

UNIVERSITY OF SOUTHERN CALIFORNIA

Center for Applied Financial Economics (CAFE)

Exchange Rate Predictability

Electronic copy available at: https://ssrn.com/abstract=2316312

July 14, 2013

Electronic copy available at: https://ssrn.com/abstract=2316312

Electronic copy available at: https://ssrn.com/abstract=2316312

Electronic copy available at: https://ssrn.com/abstract=2316312

Electronic copy available at: https://ssrn.com/abstract=2316312

Electronic copy available at: https://ssrn.com/abstract=2316312

Electronic copy available at: https://ssrn.com/abstract=2316312

Electronic copy available at: https://ssrn.com/abstract=2316312

to obtain a sequence of P h-step-ahead out-of-sample forecast errors, "ft+hjt st+h st bt ft ,

3 Which Predictors To Use?

Electronic copy available at: https://ssrn.com/abstract=2316312

Et (st+h st ) = + it+h it+h ; (1)

Electronic copy available at: https://ssrn.com/abstract=2316312

st = + (pt p t ) + "t : (2)

Electronic copy available at: https://ssrn.com/abstract=2316312

st = it+1 it+1 (yt yt ) + (mt mt ) : (3)

st = it+1 it+1 (yt yt ) + (mt mt ) + 2 (pt pt ) ; (4)

Electronic copy available at: https://ssrn.com/abstract=2316312

st = it+1 it+1 (yt yt ) + 1 (mt mt ) + 2 zt + "t : (5)

Electronic copy available at: https://ssrn.com/abstract=2316312

st = 0 + 1 (it it Et (st+1 st )) + bt bt ; (6)

Electronic copy available at: https://ssrn.com/abstract=2316312

it+1 = (1 )( + t + ytgap + qt ) + it + vt+1

it+1 = (1 )( + t + ytgap + qt ) + it + vt+1

Et st+1 st = e + eqt + e t + e ytgap + e t + eytgap + it it ; (7)

Electronic copy available at: https://ssrn.com/abstract=2316312

3.3 External Imbalance Measures

Electronic copy available at: https://ssrn.com/abstract=2316312

3.4 Commodity Prices and Other Predictors

Electronic copy available at: https://ssrn.com/abstract=2316312

3.5 What Have We Learned?

4 Which Models To Choose?

Electronic copy available at: https://ssrn.com/abstract=2316312

4.1 Single-equation, Linear Models

Electronic copy available at: https://ssrn.com/abstract=2316312

4.2 Single-equation, ECM models

Electronic copy available at: https://ssrn.com/abstract=2316312

4.3 Non-linear Models

Electronic copy available at: https://ssrn.com/abstract=2316312

4.4 Time-Varying Parameter (TVP) Models

Electronic copy available at: https://ssrn.com/abstract=2316312

4.5 Multivariate Models

Electronic copy available at: https://ssrn.com/abstract=2316312

Electronic copy available at: https://ssrn.com/abstract=2316312

4.6 Panel Models

Et (si;t+h si;t ) = h (fi;t si;t ) ;

Electronic copy available at: https://ssrn.com/abstract=2316312

5 Which Data To Use?

5.1 End-of-sample versus Filtered Data and Calibrated Parame-

Electronic copy available at: https://ssrn.com/abstract=2316312

Electronic copy available at: https://ssrn.com/abstract=2316312

5.2 Forecasted versus Realized (or Ex-post) Fundamentals

Electronic copy available at: https://ssrn.com/abstract=2316312

5.3 Countries and Samples

Electronic copy available at: https://ssrn.com/abstract=2316312