International Review of Economics and Finance: Petre Caraiani

International Review of Economics and Finance 51 (2017) 60–81
Contents lists available at ScienceDirect
International Review of Economics and Finance

journal homepage: www.elsevier.com/locate/iref
Evaluating exchange rate forecasts along time and frequency

Petre Caraiani
Institute for Economic Forecasting, Romanian Academy, Calea 13 Septembrie no. 13, Bucharest, Romania
A R T I C L E I N F O A B S T R A C T
JEL classification: A key puzzle in international macroeconomics is the proposition by Meese and Rogoff (1983) that
C53 no model can outperform the random walk in predicting the exchange rate. This paper contributes
F31 to this literature by performing an evaluation of exchange rate forecasts of reference models in
F37 time and frequency. While the literature has usually addressed the performance of exchange rate
models relative to the random walk in time only, this paper studies whether this relative per-
Keywords: formance is uniform along different frequencies or whether it is driven by certain frequencies. The
Exchange rates main finding of this paper is that the predictability of exchange rates varies along the different
Wavelets frequencies. Furthermore, the absence of the high frequency component leads to many cases in
Forecasts which the random walk is outperformed.
1. Introduction
Ever since the seminal contribution by Meese and Rogoff (1983) proposed the so-called Meese-Rogoff puzzle, i.e. that no model can
surpass the random walk in forecasting the exchange rate, there has been a continuous effort to address the puzzle via various models
and techniques. More than 40 years after this puzzle was formulated, we can still confidently state that exchange rates are basically not
forecastable, except for particular cases, see the recent review by Rossi (2013).
There is a wide body of literature that tries to outperform the random walk in forecasting the exchange rate using various methods,
and which is almost impossible to summarize. Recent research, see Rossi (2013), has found that predictability depends on several key
factors: that the predictors are the Taylor rule and/or net foreign assets, that the model is linear and that the model is small (consisting of
only a few parameters).
So far, the findings in the literature tend to suggest that the role of frequency in forecasting the exchange rates is negligible. However,
until now the research has considered the role of frequency relative only to the issue of data selection. Rossi (2013) underlines that there
are no significant differences with respect to the forecasting accuracy of daily and monthly data, but only due to using in-sample versus
out-of-sample. A potential exception to this general trend might be the study by Ferraro, Rogoff, and Rossi (2011); however, as Rossi
(2013) again underscores, the difference accounting for its superior predictability might be its use of both daily data and commodity
prices.
This paper addresses the issue of exchange rate forecastability through a new methodology, with a focus on the frequency com-
ponents. I propose the use of wavelets in the context of forecasting exchange rates. Wavelets have rarely been used in the context of
economic forecasts, notable exceptions being the studies by Michis (2014) or Rua (2011). While a more detailed presentation is
postponed to the section dedicated to wavelets, I will underline here that the key advantage of wavelets is the possibility of estimating
E-mail address: caraiani@ipe.ro.
http://dx.doi.org/10.1016/j.iref.2017.05.002
Received 21 June 2016; Received in revised form 5 May 2017; Accepted 8 May 2017
Available online 20 May 2017
1059-0560/© 2017 Elsevier Inc. All rights reserved.
P. Caraiani International Review of Economics and Finance 51 (2017) 60–81
and forecasting with the models along both time and frequency. By performing estimations and forecasts of economic models along
different frequencies, I am able to compare whether the predictability of exchange rates is influenced in any way by the frequency
dimension.
While some contributions have been made with respect to the use of wavelets in the context of forecasting, see Michis (2014) or Rua
(2011), their studies are rather limited by the fact that they do not take into account the double-side filtering property of wavelets, which
can lead to a distortion of estimation and forecasts.
At the same time, the paper is related to a growing literature pointing to the fact that, under certain conditions related to filtering the
data in an appropriate manner, wavelets can be confidently used to forecast economic time series, see Aussem, Campbell, and Murtagh
(1998), Renaud, Starck, and Murtagh (2005), Soltani, Boichu, Simard, and Canu (2000), or Zheng, Starck, Campbell, and Murtagh
(1999).
That the frequency components of exchange rate do not have uniform effects has not been a central issue in the literature until very
recently. Rabanal and Rubio-Ramrez (2015) have shown that the low frequency component of the real exchange rate is more important
in shaping the dynamics of unfiltered real exchange rates than the usually analyzed business cycle component.
Though a bit less strictly related, the paper also indirectly touches on how the structural relationships and/or causality relationships
look differently when examining the shocks affecting the variables in question, see Caraiani (2016), for a study on the causality between
money, on one hand, and output and inflation on the other hand. It nevertheless contributes to a growing literature showing that
aggregate macroeconomic relationships can be broken using structural decompositions like above, or wavelet decompositions, see
Gallegati, Gallegati, Ramsey, and Semmler (2011).
The main contributions of this paper are as follows. I propose and discuss alternative methodologies to assess forecasting accuracy
through the use of wavelets filtering in time and frequency. Second, I perform a wavelet filtering that addresses the double-sided nature
of wavelets, a key issue in forecasting. Third, I estimate and forecast with various models along the different frequencies, comparing
whether the forecasting accuracy of exchange rates is influenced in any way by the frequency dimension. Fourth, I contribute to the new
results in the literature regarding the role of frequency components in shaping the dynamics of the exchange rate, see Rabanal and
Rubio-Ramrez (2015) for a recent result. Finally, the comparison is conducted across a wide range of countries, data series, models, and
tests for forecast accuracy, further strengthening the main findings.
The key finding of the paper is that, in contrast to what is found in previous literature on exchange rate forecasting, there are
evidences that the frequency dimension does matter for the accuracy of exchange rate forecasts. Not only does each frequency con-
sidered alter forecasting accuracy, there are also stark differences between the high frequency components and the low frequency
components. When the high frequency component is removed, the random walk is outperformed for several countries and models, while
removing the low frequency component drastically weakens exchange rate predictability.
2. Models
Following Rossi (2013), I focus on the most widely-used models in the literature, selecting those for which quarterly data was
available. There are several reasons to follow this approach. First of all, it is one of the best summaries of the literature to date. Second, in
contrast to most previous studies, it includes only the best-performing and widely-used models in its comparison. Lastly, it allows a focus
on the role of frequency for a dataset and models already used in a previous study.
I focus on two classes of models, univariate linear models as well as multivariate linear models. I exclude nonlinear models, since, as
Rossi (2013) suggests, their performance in the literature to date is rather modest.
Dating back to Fisher (1896), the Uncovered Interest Parity (UIRP, hereafter) model has been widely used in the literature. I use a
standard specification:

Et ðstþh st Þ ¼ α þ β itþh itþh (1)
Here st stands for the logarithm of the nominal bilateral exchange rate St , i.e., st ¼ lnðSt Þ; it is the domestic nominal interest rate,
while it is the foreign nominal interest rate; the parameter α ¼ 0 while β ¼ 1, while h stands for the horizon.
Unfortunately, to date there are scarce evidences favoring the UIRP model, with some limited positive results in Clark and West
(2006) at short horizons and Molodtsova and Papell (2009) for a sample of countries.
An equally frequently used model in forecasting the exchange rate is based on Purchasing Power Parity (PPP, hereafter). In its
standard form, the PPP model can be written as:

st ¼ α þ β pt p*t þ εt (2)
pt
Here pt stands for the logarithm of the domestic commodity price CP index, is the logarithm of the foreign commodity price index,
the parameter α ¼ 0 and β ¼ 1, while h stands for the horizon.
The evidence does not favor the PPP model either, see the reference study by Cheung, Chinn, and Pascual (2005), although some
limited evidence has been found for this model at long horizons, which is not, however, significantly better statistically.
A third single equation model to be used is the monetary model. This model dates back to the contributions by Frenkel (1976) and
Mussa (1976) and builds on a very basic real demand for money, which can be written as:
61
mt pt ¼ ηitþ1 þ ϕyt (3)

Here pt stands for the logarithm of the domestic commodity price CP index, mt is the logarithm of nominal money, and mt is the
logarithm of real output. Assuming the same relationship holds for the foreign economy, and denoting with a star the corresponding
variables in the foreign economy, we obtain the flexible price version of the monetary model, as follows:

st ¼ η itþ1 itþ1 ϕ yt yt þ mt mt (4)
Summarizing the evidences regarding the empirical evidences in the favor of the monetary model, Rossi (2013) finds them to be
rather scarce, with some positive in-sample evidences. The literature points instead to clearly superior forecasts obtained with the
random walk model, see Meese and Rogoff (1983), confirmed later by Chinn and Meese (1995) for short horizons. More recent research
also dismisses the longer-run forecasting performance, see Rossi (2005c).
A more recent single equation model with promising results is the Taylor model, see Engel and West (2005, 2006) or Molodtsova and
Papell (2009). As the name suggests, this is based on Taylor's well-known monetary policy rule, see Taylor (1993).
Molodtsova and Papell (2009) modify the standard Taylor rule to account for the fact that in an actual economy, monetary policy also
depends on the real exchange, while interest rate changes are sluggish. The specification they base their model on is given by:

itþ1 ¼ ð1 ρÞ μ þ λπ t þ γygap
t þ δqt þ ρit þ vtþ1 (5)
where π t is the inflation rate while ygap

tis the output gap. When moving to an open economy setting, and considering the UIRP, the Taylor
rule can be written as (seeMolodtsova and
Papellgap(2009)):

Et stþ1 st ¼ μ~ þ ~λ π t π t þ ~γ ygap
t yt (6)
This specification is known as the ”symmetric homogeneous” Taylor rule, with the lagged interest rates excluded.
In contrast to previous models, both the in-sample and out-of-sample evidences tend to be positive, see Molodtsova and Papell
(2009). More recent research has taken into account the changing nature of monetary policy during and after the last financial crisis by
including various financial factors within the Taylor rule, see Molodtsova and Papell (2013).
There are two widely-used versions of the monetary model the baseline specification in equation (3), as well as an error correction
model (ECM, hereafter), as proposed by Mark (1995). The specific feature of this model is the assumption of a long-run relationship
between the exchange rate st and the fundamentals ft . The model is specified below:

Et stþh st ¼ α þ β mt mt ϕ yt yt st (7)
By construction, the monetary ECM should have a good performance in the long run. Some evidences favoring its forecasting ability
in the long run were provided by Mark (1995), but Cheung et al. (2005) and Alquist and Chinn (2008) failed to find a better forecasting
performance relative to the random walk model. However, based on the very same specification as for the ECM model, Rossi (2005a)
found some predictive ability for this model.
The selection of univariate models is basically similar to the selection in Rossi (2013), except that I exclude the portfolio balance
model since the available data for this model is at quarterly frequency while the focus of this paper is on performing a wavelet
decomposition, and I prefer focusing on the data at the same frequency, i.e. monthly data.
In terms of multivariate models, I again follow Rossi (2013) and select the same models as in her analysis. Two such models were
selected, the panel equilibrium correction model (panel ECM, hereafter) and the Bayesian Model Averaging (BMA, hereafter).
I use the specification for the panel ECM as suggested by the literature, see Mark and Sul (2001), Cerra and Saxena (2010) or Rapach
and Wohar (2006), namely:

Et ðstþh st Þ ¼ βh fi;t γii;t (8)
The evidences from the literature tend to be positive with respect to the panel ECM, see Mark and Sul (2001) or Cerra and Saxena
(2010).
A final model used is the Bayesian moving average (BMA, hereafter). Following Wright (2008), it has been shown in the literature
that we can combine many predictors, via Bayesian averaging with weights estimated by posterior probabilities. Though Wright (2008)
found that for most currencies, BMA is able to outperform the random walk model, the differences are minor in terms of statistical
significance.
3. Data selection
Comparing many models over the same data for several countries has limited the possibility of data selection. Since my intention is to
build on the work by Rossi (2013), I use the dataset she provided. The dataset consists in the exchange rates relative to the United States,
as well as key economic fundamentals for several countries: Australia, Austria, Belgium, Canada, Denmark, Finland, France, Germany,
Greece, Ireland, Italy, Japan, New Zealand, Portugal, Spain, Sweden, Switzerland, and the United Kingdom.
I focus on monthly data, in order to consistently use the same wavelet filtering method across data and models. This does not
however limit the number of models, since I exclude only the portfolio model from my analysis. The selected monthly data consists of
62
nominal interest rates, three-month Treasury Bills, five-year Treasury Bonds, industrial production, CPI and money stocks.
Given the changes in the statistics of various countries (due to the way data are reported or collected), the samples across the various
countries differ. Appendix A presents the available samples for data series across the countries included in the analysis.
Following the arguments in Rossi (2013), the series are not seasonally adjusted at first. The seasonal adjustment process was done
using a uniform procedure along all series, namely a one-sided moving average with backward, equal weights.
One issue not addressed by Rossi (2013) is that many of the variables, including the exchange rates, start before the 1970s. This
implies using data before the post-Bretton Woods period. However, on one hand, the present paper builds on the earlier work by Rossi
(2013) in order to contrast the results with the ones by Rossi, and, on the other hand, due to the nature of the wavelet filtering (requiring
samples that have a sample size of the power of 2) the actual samples used may very well not cover observations before 1970.
4. Wavelet decomposition
In this section, I discuss the wavelet decomposition used throughout the paper. I present the wavelet transform used and details
regarding its practical implementation, as well the issue of double-side filtering.
In essence, this paper uses the discrete wavelet transform in order to estimate time series models with traditional econometric
approaches. The alternative continuous wavelet transform is not a good fit for this kind of econometric exercise.
In the following paragraphs, I detail the theoretical basis of wavelets, before turning to their empirical implementation. The pre-
sentation here follows the more detailed descriptions in Gencay, Selcuk, and Whitcher (2001) and Percival and Walden (2000).
Economists are much more familiar with the Fourier transform. As Gencay et al. (2001) show, the Fourier transform can be seen as a
combination of sine and cosine functions and, in the end, as resulting from a component at different frequencies. However, the major
disadvantage of the Fourier transform is its inability to deal with the time dimension. An improvement over the Fourier transform that
tries to address this shortcoming is the Short Time Fourier Transform (STFT, hereafter). Unfortunately, as Gencay et al. (2001) observed,
this approach cannot capture the movements within the sliding window, which becomes critical for the successful application of this
approach.
In order to address this, the wavelets approach was proposed. This approach can be used to filter a series in both time and frequency,
thus addressing the shortcomings of both Fourier transform and STFT.
It is well known that when working with a model that forecasts t þ h steps ahead, it is essential not to use the information between t
and t þ h in the estimation. This issue affects double-side filters such as the Hodrick-Prescott filter, wavelets, etc. I propose and discuss
two approaches: one based on the Maximal Overlap Discrete Wavelet Transform which requires however an additional way to deal with
the use of forward information in estimating and forecasting the models discussed, and a second approach that deal directly with this,
based on the Haar redundant wavelet transform.
4.1. Maximal overlap discrete wavelet transform
The discrete wavelet transform has the following properties, summation to zero:
X
L1
hl ¼ 0 (9)
l¼0
and unit energy:

XL1
h2l ¼ 1 (10)
l¼0
Here hl ¼ h0 ; h1 ; …; hL1 is a discrete wavelet filter with a finite length. The wavelet hl is known to be orthogonal to its shifts:
X
L1
hl hlþ2n ¼ 0 (11)
l¼0
for all nonzero integers n.

I focus on the Maximal Overlap Discrete Wavelet Transform (MODWT, hereafter). This has been successfully used in the economics
literature before, see for example Michis (2014) or Rua (2011), who used it in forecasting, or Gallegati et al. (2011) or Caraiani (2015),
who estimated structural models. MODWT has certain advantages over the baseline discrete wavelet transforms: we can use samples
that are not multiples of 2, and at the same time, MODWT is invariate to shifting.
Gencay et al. (2001) provide a thorough discussion of the advantages of the MODWT. Let us start from a time series xt of length N.
Using MODWT, we can obtain the vector of coefficients w ~ of length (J þ 1)N by using:
w ~
~ ¼ Wx (12)
~ is a (J þ 1)NxN matrix which determines the MODWT. The coefficients of the MODWT, can be written in J þ 1 vectors as
Here W
follows:
63
T
~¼ w
w f1 ; …; w
fJ ; veJ (13)
where wej is a vector of length N=2j with the wavelets coefficients corresponding to changes on a scale of length λj ¼ 2j1 , while veJ is a
vector with a length N=2J , consists in values of scaled coefficients corresponding to averages on a scale of length 2J ¼ 2λj . Gencay et al.
(2001) underline that MODWT wavelets and scaling coefficients can be transformed into the wavelets and scaling coefficients corre-
sponding to DWT using:
~ j;2j ðjþ1Þ1
wj;t ¼ 2j=2 w (14)
for t ¼ 0; …; N=2j 1, for the wavelet coefficients, and
vJ;t ¼ 2J=2 ~vJ;2J ðjþ1Þ1 (15)
for t ¼ 0; …; N=2 1 for the scaling coefficients. Of course, this applies for a series of dyadic length N ¼ 2J.
J
As we can generally do for the discrete wavelet transform, we can also apply a multi-resolution analysis for the MODWT. By doing
this we can decompose a given series into the sum of its details and the smooth component as follows:
X
Jþ1
xt ¼ d~j;t (16)
j¼1
~j;t is the component t of d

for t ¼ 0; ; N 1. Here d ~j , or the detail at level j, defined by:
d~j ¼ W
~ Tj w
~j (17)
for j ¼ 1; ; J. The smooth component can be obtained by using:
X
Jþ1
~sJ;t ¼ d~k;t (18)
k¼jþ1
for t ¼ 0; ; N 1.
I use a practical implementation of the MODWT that has been used before in the literature: the Daubechies least asymmetric wavelet
filter (henceforth, LA) with a length of 8, following Gallegati et al. (2011) and similar papers. I use six scales, which is perfectly justified
for monthly data, since Gallegati et al. (2011) used four scales for quarterly data. I denote the details of the decomposition by W, W1 to
W6 and the smooth component by S6. W1 reveals the dynamics at 2–4 months; the W2 scale uncovers the dynamics at 4–8 months. The
scales W3 and W4 give the results between 8 and 16 months for W3 and between 16 and 32 months for W4. Finally, the scales W5 and
W6 show between 32 and 64 months for W5 and between 64 and 128 months for W6, while the smooth component V6 details the
dynamics at low frequencies, that is, for periods longer than 128 months (basically, 10 years).
I designed a practical way to circumvent this issue when filtering with wavelets. Since the window on which the models are
recursively estimated represents half of the sample, I split the series into two halves. By applying the wavelet decomposition for the first
half, I obtain the details for this sub-sample. To filter the observations in the second half, I recursively increase the sample, adding one
observation at a time starting from the original half and filtering the new sub-sample each time. This is done up to the end of the sample.
Each time I add an observation and filter with wavelets, I also add an observation to the already-filtered details. This way, when
estimating the models recursively up to time t, I do not use information from t þ h steps ahead. Instead, for each observation t in the
second half of each data series, I use only information from the first observation to observation t.
4.2. The redundant Haar wavelet transform
In this section, I detail the redundant wavelet transform used to forecast time series. The presentation follows the contributions by
Aussem et al. (1998) and Zheng et al. (1999).
The discrete wavelet transform (DWT) is known to suffer from the inability to relate the information in time with the information
provided by the decomposition along different scales. Another deficiency is the lack of shift invariance such that removing a few
observations leads to a different decomposition.
The MODWT addressed these deficiencies since it is a redundant transform too. However, since when doing predictions, the derived
wavelet coefficients must not be based on future information, as it is implied by the MODWT algorithm, I proposed in the previous
section a practical implementation that overcomes this issue. At the same time, the recursive method is far from an ideal approach too.
The MODWT is based on periodic or reflecting boundary conditions which make the wavelet decomposition rely on an extrapolation.
What is important is that the recursive method proposed earlier might be sensitive to introducing new observations at the end-points.
To overcome these issues, I use a redundant Haar Wavelet Transform, which has the advantage of making the time-scale decom-
position based exclusively on the previous data-points.
I first present the algorithm for general redundant discrete wavelet transform, which is also known as the a trous wavelet transform.
64
The data to be filtered is decomposed into the wavelet coefficients and a low-pass filter of the initial data series. We start from an initial
series denoted by c0 ðkÞ, to distinguish it from the smooth filter at scale j, denoted by cj ðkÞ. We define it as being given by the scalar
product at samples k of the function f ðxÞ with a scaling function ϕðxÞ.
c0 ðkÞ ¼ < f ðxÞ; ϕðx kÞ > (19)

The scalar function ϕ is chosen in such a manner that it satisfies the so-called dilation equation:
1 x X
ϕ ¼ hðlÞϕðx lÞ (20)
2 2 l
Here, h is defined as a low-pass filter that corresponding to the scaling function ϕx . Given these relations, we can obtain the smoothed
component at resolution j for the observation k as:
1 ðx kÞ
cj ðkÞ ¼ < f ðxÞ; ϕ > (21)
2j 2j
Considering two consecutive resolutions, we might denote the difference between them by wj . Then, we could write:
wj ðkÞ ¼ cj1 ðkÞ cj ðkÞ (22)
But this can be alternatively written as:
1 ðx kÞ
wj ðkÞ ¼ < f ðxÞ; ψ > (23)
2j 2j
trous algorithm. Here ψ is a wavelet function defined as follows:
This is the discrete wavelet transform based on the a
1 x 1 x
ψ ¼ ϕðxÞ ϕ (24)
2 2 2 2
Following this algorithm, we obtain in the end a decomposition of the original series c0 . The latter can be written as:
X
p
c0 ðkÞ ¼ cp þ wj ðkÞ (25)
j¼1
This is not enough however to ensure that when filtering the information at the moment k, we don't use the information from the
future k þ i, but only the information provided by the past values, k i, with i 0. To ensure this, we further use the Haar wavelet
transform, a non-decimated version of it to be more precise. In this approach, the low-pass filter h from the a trous algorithm given by:
ð1=16; 1=4; 3=8; 1=4; 1=16Þ is replaced with the simpler filter ð1=2; 1=2Þ that has the essential advantage of being non-symmetric. This
implies that filtering with the Haar redundant algorithm consists in using only past information when filtering the observation c0 ðkÞ of
the original series.
4.3. Forecast evaluation of the different frequencies
The purpose of this paper is to evaluate the forecasts accuracy across time and frequency, which can be done in many ways. The most
direct way is to estimate and forecast the models along the different details, which is how it has been usually done before, see Michis
(2014) or Rua (2011). However, given that the details of the decomposition have a pronounced frequency pattern, this does not seem the
best way to uncover the impact of different frequencies. Accordingly, I instead evaluate the marginal impact of the different details by
constructing counterfactual series as follows:
A series X can be represented as:
X
6
Xt ¼ DJt ð :; iÞ þ SJt ð :; 1Þ (26)
i¼1
where DJt'i are the details while SJt is the smooth component. To quantify the impact of different frequencies, the construct counter-
factual series are built by taking away one detail at a time.
Since we are interested in a time-scale evaluation of all the model presented above, all the series used are filtered using the same
methodology, including here both the predictors and the dependent variables. Although the paper might have benefited by doing
multivariate decompositions that account for the covariance among the variables, we use a rather standard framework, although we do
underline that future studies might incorporate this too.
5. Results
In this section, I compare the forecasting performance of the selected univariate and multivariate models at two horizons: a short
65
horizon h1 equal to 1 month and a long horizon h2 equal to 48 months, that is, 4 years. I use as benchmark the random walk without drift.
The models are recursively estimated starting from a rolling window with a size equal in dimension to half of the size of the actual
series, which differ in length. The forecast assessment is done both in-sample and out-of-sample. I focus on four key statistics, the
Granger-causality test (GC, hereafter) adjusted robust to instabilities (see Rossi (2005b)), the Root Mean Square Error Relative to the
random walk (RSMSER, hereafter), as well as two traditional tests for the power of out-of-sample forecasts, namely Diebold Mariano
(DM, hereafter) and Clark-West (CW, hereafter).
The results of the forecast assessment are reported in Appendix C. I estimate each model across each country in the sample for the two
forecasting horizons, h1 and h2. Each estimation was performed along the five different counterfactual series (that is, one detail at a time
being taken off). I report the probability value associated to the GC test, RMSER, and the probability values associated to DM (pvDM in
the tables) as well as CW tests (pvCW in the tables). The significance threshold is taken at 5 percent, such that p-values lower than 0.05
denote significance. N.a. marks cases where the estimation could not be performed, including cases where the series was too small to
perform a wavelet decomposition, usually for less than 30 observations.
Given the amount of results, I don't report the results based on applying the recursive wavelet filtering, but present only the results
based on the redundant Haar transform. One reason is the limited space, while a second reason, more significant, is related to the
limitations of the proposed recursive MODWT filtering, as discussed at the beginning of section 4.2. However, it should be mentioned
that, qualitatively, the alternative procedure using a recursive filtering leads to similar results as the procedure based on redundant Haar
wavelet filtering.
5.1. Model performances
I start by discussing the results from forecasting with the models along the countries, models and counterfactual series as presented in
Appendix C.
In terms of in-sample predictability, we can see a very strong impact of eliminating the high frequency component W1, leading to
strong in-sample predictability. As we move along the different counterfactual series, the in-sample predictablity becomes weaker. The
monetary ECM and the Taylor models seem least affected by the frequency component removed. Even when detail W5 is removed, there
is some in-sample predictability for these models.
When moving to out-of-sample predictability, some of the patterns found for the in-sample predictability remain, while other
findings are different. First of all, there are similar results in terms of eliminating the high frequency component W1. For this case, the
results indicate a tendency across most models (except UIRP, the monetary ECM and the BMA models), for some countries, to out-
perform the random walk in a statistical significant manner (confirmed by either DM or CW statistics). This is quite interesting, since
usually models like PPP or the monetary model have weak predictability. The strongest evidence comes from the monetary panel. When
moving to the series without lower frequency details, i.e. W2 to W4, the evidences of predictability are clearly weaker. The counter-
factual series without the low frequency detail W5 indicate that the predictability power of the models is almost non-existent, except a
few particular cases (especially the Taylor rule and the PPP model).
5.2. Further statistics
I also summarize the empirical findings by using scatterplots for both the in-sample and out-of-sample forecasts. The results are shown in
Appendix B. For each counterfactual series, there is scatterplot presenting the CW statistic as well as a scatterplot presenting the GC statistic.
In the scatterplots, each point stands for predictors for each country and each model. Again, a p-value smaller than 0.05 denotes statistical
significance at a level of 5 percent. On each scatterplot, the x-axis stands for the short horizon, while the y-axis shows the long horizon.
The results in Rossi (2013) summarize very well the predictive ability of the Taylor rule and monetary panel for both in-sample and
out-of-sample for the series which are not filtered with wavelets.
What can we learn from our forecasting exercises along time and frequency? When the high frequency detailW1 is eliminated, as
seen from Figs. B.1 and B.2, we can see that most of the models are significant predictors (for some countries) at in-sample, generally at
both short-run and long-run horizons. For the out-of-sample performance, some models (especially PPP and the monetary ECM) tend to
be significant especially at short-run horizons (and we can include here most models), while the Taylor model and monetary ECM and
the monetary model are significant at long run horizons.
The scatterplots for each counterfactual series indicate significant variance, suggesting a strong impact from the frequency com-
ponent. However, we can notice a pattern: as we move along the counterfactual series where details from lower frequencies are taken
out, fewer models have statistically significant predictive power. For example, the counterfactual series without the low frequency
component W5 show that just a few models have statistically significant predictive power. For in-sample, this happens for the monetary
ECM and the monetary model at short horizons, as well as the monetary ECM, Taylor rule and monetary model at long horizons. For out-
of-sample, the Taylor rule and the monetary panel have statistically significant predictive power.
5.3. A comparison with the results in the literature
The research here builds on the earlier work by Rossi (2013). Due to limited space, and since they are already available in that earlier
work, I do not present here the results for series not filtered with wavelets, but direct the reader to the original article. Instead, I compare
the results in this forecasting exercise focusing on the role of frequency with those in the paper by Rossi (2013).
The main finding in Rossi (2013) is that while several models display in-sample predictability, only a few of them also display out-of-
66
sample predictability, namely the Taylor model at short horizons and the monetary panel model at long horizons.
In contrast, I show on the very same data and with identical models that the frequency component can drastically alter the results.
For example, removing the high frequency component W1 leads to most of the models having statistical significant in-sample predictive
power. For out-of-sample, we see significant short-run prediction power for most models except the Taylor rule, while for the Taylor rule
there is predictive power at long horizons, along with the monetary ECM and monetary model.
Interestingly, for out-of-sample, the Taylor rule seems to remain robust along the different frequencies, having significant predictive
power mostly at long-run horizons. I also find that the monetary panel, the monetary model and the monetary ECM possess predictive
power along the different frequencies.
However, there are models which also display out-of-sample forecasting ability at different frequencies, like UIRP, PPP and the
monetary model at short horizons when excluding the high frequency detail W1, or the PPP or the monetary panel when excluding the
low frequency detail W5.
6. Conclusion
This paper has asked whether the frequency dimension of the predictors matters for exchange rate predictability. The results show
clear evidence that the frequency components do matter. Furthermore, the different frequency components do not affect the predict-
ability in a uniform matter: the high-frequency components tend to negatively influence the accuracy of forecasts.
The results are robust to the way they are filtered with wavelets, as two methods are applied: a recursive approach based on the
MOWDT, and a direct filtering with the Haar redundant wavelet transform. Both methods ensure that the current filtered values of a
series do not take into account the future values of the same series, a key issue when doing forecasting.
Resuming earlier work, Rossi (2013) shows that detrending, filtering and seasonal adjustment significantly determine predictability.
This study, working on similar data as in Rossi (2013), and performing an analysis across time and frequency, indicates that the fre-
quency components have non-negligible effects, depending on the frequency scale.
The frequency component affects the predictive performance of various models at both short horizons and long horizons. The effects
are not uniform across models, type of forecasting (in-sample or out-of-sample) and horizon of forecasting. However, there is a clear
tendency to see that when the lower-frequency components are removed (e.g. W5), the predictive power of models decreases, while
when high-frequency components are removed (especially W1), the predictive performance increases drastically across all models. This
finding is meaningful from the economic point of view, since the high frequency movements are hardest to predict, while lower fre-
quency components tend to move more in line with various fundamentals.
Acknowledgement
This work was supported by CNCSIS - UEFISCSU, project number PNII IDEI nr. 42/2.09.2013.
Appendix A. Data series samples across countries
Table 1
Models in monthly data.
Country Exchange rate differential Interest rate differential CPI differential Money differential Output differential
Start End Obs Start End Obs Start End Obs Start End Obs Start End Obs
Australia 1957:02 2011:06 653 1969:08 2011:06 503 NaN NaN NaN 1976:02 2011:06 425 1958:02 2007:09 596
Belgium 1957:02 1998:12 503 1957:02 1999:01 504 1958:02 2011:06 641 2003:01 2011:06 102 1958:02 2011:03 638
Canada 1957:02 2011:06 653 1975:02 2011:06 437 1958:02 2011:06 641 1976:03 2011:06 424 1958:02 2011:05 640
Denmark 1957:02 2011:06 653 1972:02 2011:06 469 1981:01 2011:06 366 1992:01 2011:06 234 1958:02 2011:05 640
Finland 1957:02 1998:12 503 1978:01 2011:06 402 1961:01 2011:06 606 1961:01 2011:06 606 1958:02 2011:06 641
France 1957:02 1998:12 503 1964:02 1999:03 418 1991:01 2011:06 246 1978:12 2011:06 391 1958:02 2011:05 640
Germany 1957:02 1998:12 503 1960:02 2011:06 617 1958:02 2011:06 641 1974:01 2011:06 450 1959:01 2011:05 629
Greece 1957:02 2000:12 527 1998:02 1999:10 021 1960:01 2011:06 618 1981:01 2011:06 366 1958:02 2011:05 640
Ireland 1957:02 1998:12 503 1972:04 2011:06 450 1970:0 1 2011:06 498 2000:01 2011:06 138 1977:01 2011:05 413
Italy 1957:02 1998:12 503 1971:02 2011:06 485 1958:02 2011:06 641 1981:01 2011:06 366 1958:02 2011:05 640
Japan 1957:02 2011:06 653 1957:02 2011:06 653 2001:01 2011:06 126 1961:01 2011:06 606 1958:02 2011:06 641
New Zealand 1957:02 2011:06 653 1985:02 2011:06 317 NaN NaN NaN 1978:03 2011:06 400 1988:01 2007:12 240
Portugal 1957:02 1998:12 503 1983:02 2000:03 206 1961:01 2011:06 606 1998:09 2011:06 154 1958:02 2006:04 579
Spain 1957:02 1998:12 503 1974:02 2011:06 449 1958:02 2011:06 641 1998:09 2011:06 154 1962:01 2011:05 593
Sweden 1957:02 2011:06 653 1966:01 2011:06 546 1961:01 2011:06 606 1999:01 2011:06 150 1958:02 2011:05 640
Switzerland 1957:02 2011:06 653 1975:10 2011:06 429 1958:02 2011:06 641 1985:12 2011:06 307 1996:01 2007:12 144
U.K. 1957:02 2011:06 653 1972:02 2011:06 473 1989:01 2011:06 270 1987:09 2011:06 286 1958:02 2011:05 640
67
Appendix B. In and out-of-sample performance across details
Fig. B.1. Scatterplot GC: The Series without W1.
Fig. B.2. Scatterplot: The Series without W1.
68
69
70
Appendix C. The performance of the models
Table 1a
The performance of single equation models (without W1).
Country 1.UIRP model 2.PPP Mode 3.Monetary Model
h ¼ 1 month h ¼ 1 month h ¼ 1 month
GC RMSER pvDM pvCW GC RMSER pvDM pvCW GC RMSER pvDM pvCW
Aus 0.43 1.00 0.50 0.29 NaN NaN NaN NaN 0.00 0.98 0.46 0.01
Bel 0.03 1.01 0.54 0.67 0.95 1.01 0.56 0.83 NaN NaN NaN NaN
Can 0.20 1.00 0.50 0.30 0.00 0.97 0.43 0.00 0.81 1.01 0.53 0.47
Dnk 0.20 1.01 0.53 0.55 0.00 1.00 0.50 0.18 0.68 1.03 0.55 0.30
Fin 0.29 1.02 0.56 0.78 0.06 1.01 0.55 0.87 0.39 1.03 0.56 0.77
Fra 0.17 1.01 0.53 0.64 0.12 1.05 0.60 0.83 0.11 0.97 0.45 0.03
Ger 0.18 1.01 0.54 0.67 0.38 1.01 0.53 0.55 0.00 1.01 0.51 0.06
Gre NaN NaN NaN NaN 0.52 0.98 0.48 0.02 0.00 0.98 0.47 0.00
Irl 0.91 1.01 0.53 0.48 0.00 1.03 0.56 0.72 NaN NaN NaN NaN
Ita 0.27 1.00 0.52 0.55 0.00 1.01 0.53 0.38 0.00 1.06 0.55 0.24
Jpn 0.44 1.00 0.50 0.21 0.89 1.03 0.54 0.25 0.03 1.03 0.56 0.89
Nzl 0.89 1.00 0.50 0.37 NaN NaN NaN NaN 0.00 0.96 0.43 0.01
Por 0.89 0.99 0.47 0.14 0.00 1.01 0.53 0.38 NaN NaN NaN NaN
Spa 0.38 1.01 0.55 0.79 0.25 1.01 0.54 0.57 NaN NaN NaN NaN
Swe 0.71 1.01 0.56 0.89 0.03 1.00 0.50 0.13 0.94 1.09 0.60 0.67
Swi 0.61 1.00 0.48 0.17 0.32 1.00 0.50 0.04 0.83 1.02 0.54 0.52
Uk 0.36 1.01 0.55 0.86 0.10 1.01 0.55 0.74 0.23 1.01 0.55 0.45
h ¼ 4 years h ¼ 4 years h ¼ 4 years
Aus 0.12 1.00 0.49 0.23 NaN NaN NaN NaN 0.00 0.98 0.46 0.01
Bel 0.07 1.00 0.49 0.14 0.17 0.99 0.48 0.07 NaN NaN NaN NaN
Can 0.31 1.00 0.50 0.29 0.22 1.00 0.51 0.22 0.81 1.01 0.53 0.47
Dnk 0.23 1.00 0.52 0.60 0.22 1.01 0.52 0.21 0.68 1.03 0.55 0.30
Fin 0.00 1.00 0.50 0.18 0.26 1.02 0.59 0.86 0.39 1.03 0.56 0.77
Fra 0.90 1.02 0.54 0.32 0.71 1.09 0.65 0.60 0.11 0.97 0.45 0.03
Ger 0.07 1.01 0.52 0.26 0.01 0.97 0.44 0.01 0.00 1.01 0.51 0.06
Gre NaN NaN NaN NaN 0.00 0.97 0.47 0.00 0.00 0.98 0.47 0.00
Irl 0.00 1.02 0.52 0.24 0.71 1.03 0.59 0.84 NaN NaN NaN NaN
Ita 0.11 1.04 0.59 0.58 0.23 1.01 0.52 0.28 0.00 1.06 0.55 0.24
Jpn 0.60 1.01 0.53 0.47 0.89 1.01 0.52 0.28 0.03 1.03 0.56 0.89
Nzl 0.71 1.00 0.49 0.23 NaN NaN NaN NaN 0.00 0.96 0.43 0.01
Por 0.37 1.03 0.56 0.63 0.10 1.05 0.62 0.94 NaN NaN NaN NaN
Spa 0.00 1.03 0.53 0.11 0.43 1.03 0.56 0.33 NaN NaN NaN NaN
Swe 0.05 1.01 0.56 0.89 0.45 1.02 0.57 0.81 0.94 1.09 0.60 0.67
Swi 0.93 1.00 0.49 0.24 0.42 1.00 0.48 0.15 0.83 1.02 0.54 0.52
Uk 0.78 1.01 0.54 0.72 0.00 1.02 0.55 0.11 0.23 1.01 0.55 0.45
71
Table 2a
Country 1.Monetary ECM model 2.Taylor Model
h ¼ 1 month h ¼ 1 month
GC RMSER pvDM pvCW GC RMSER pvDM pvCW
Australia 0.31 1.02 0.60 0.99 NaN NaN NaN NaN

Belgium NaN NaN NaN NaN 0.35 1.70 0.49 0.84
Canada 0.66 1.01 0.55 0.87 0.00 1.35 0.53 0.78
Denmark 0.03 0.99 0.43 0.04 0.00 1.10 0.54 0.78
Finland 0.07 1.00 0.49 0.20 0.00 1.04 0.47 0.73
France 0.03 1.01 0.54 0.37 0.20 2.04 0.72 0.28
Germany 0.16 1.00 0.51 0.29 0.00 1.67 0.51 0.83
Greece 0.50 1.01 0.56 0.79 0.00 0.65 0.53 0.15
Ireland NaN NaN NaN NaN 0.00 1.78 0.53 0.84
Italy 0.24 1.02 0.56 0.67 0.00 0.93 0.46 0.06
Japan 0.12 1.01 0.53 0.51 0.00 1.84 0.50 0.81
New Zealand 0.53 1.02 0.62 0.97 NaN NaN NaN NaN
Portugal NaN NaN NaN NaN 0.00 0.74 0.47 0.15
Spain NaN NaN NaN NaN 0.00 1.73 0.52 0.83
Sweden 0.06 1.00 0.51 0.39 0.31 1.05 0.55 0.09
Switzerland 0.76 1.02 0.63 0.99 0.00 2.16 0.45 0.84
U.K. 0.03 1.00 0.54 0.54 0.52 1.14 0.51 0.64
h ¼ 4 years h ¼ 4 years

Canada 0.12 1.01 0.54 0.58 0.00 1.35 0.53 0.78
Denmark 0.94 1.01 0.57 0.60 0.00 1.10 0.54 0.78
Finland 0.36 1.05 0.59 0.10 0.00 1.04 0.47 0.73
France 0.69 1.01 0.56 0.75 NaN NaN NaN NaN
Germany 0.80 1.01 0.58 0.87 0.00 1.67 0.51 0.83
Greece 0.86 1.00 0.50 0.16 0.00 0.65 0.53 0.15
Italy 0.95 1.02 0.53 0.14 0.00 0.93 0.46 0.06
Japan 0.76 1.00 0.50 0.30 NaN NaN NaN NaN
Sweden 0.16 1.04 0.60 0.76 0.31 1.05 0.55 0.09
Switzerland 0.44 1.00 0.52 0.30 0.00 2.16 0.45 0.84
U.K. 0.31 1.02 0.54 0.47 0.52 1.14 0.51 0.64
Table 3a
The performance of multiple equations models(without W1).
Country 1.Monetary Panel model 2.BMA Model
RMSER pvDM pvCW RMSER pvDM pvCW
Australia 0.91 0.33 0.00 1.00 0.51 0.15

Belgium NaN NaN NaN NaN NaN NaN
Canada 0.90 0.38 0.00 1.01 0.52 0.19
Denmark 0.90 0.36 0.00 1.01 0.56 0.82
Finland 0.88 0.31 0.00 1.01 0.54 0.55
France 0.91 0.33 0.00 1.00 0.50 0.13
Germany 0.89 0.33 0.00 1.00 0.50 0.22
Greece NaN NaN NaN 1.02 0.53 0.39
Ireland NaN NaN NaN NaN NaN NaN
Italy 0.93 0.38 0.00 1.03 0.55 0.33
Japan 0.89 0.35 0.00 1.01 0.52 0.54
New Zealand 0.89 0.31 0.00 1.01 0.51 0.06
Portugal NaN NaN NaN NaN NaN NaN
Spain NaN NaN NaN NaN NaN NaN
Sweden 0.88 0.38 0.00 1.01 0.57 0.83
Switzerland 0.90 0.34 0.00 1.04 0.58 0.81
U.K. 0.88 0.41 0.00 1.01 0.54 0.64
72
Table 3a (continued )
Australia NaN NaN NaN 1.45 0.77 1.00

Canada NaN NaN NaN 1.38 0.69 0.27
Denmark NaN NaN NaN 1.83 0.73 0.03
Finland NaN NaN NaN 1.75 0.72 0.00
France 0.89 0.41 0.00 1.99 0.99 1.00
Germany 0.90 0.38 0.00 1.91 0.72 1.00
Greece 0.91 0.37 0.00 0.83 0.22 0.00
Ireland 0.89 0.32 0.00 NaN NaN NaN
Italy 0.91 0.34 0.00 1.38 0.81 0.00
Japan 0.90 0.34 0.00 1.46 0.63 0.97
New Zealand 0.91 0.36 0.00 1.80 0.74 0.93
Portugal 0.89 0.35 0.00 NaN NaN NaN
Spain 0.86 0.36 0.00 NaN NaN NaN
Sweden 0.89 0.29 0.00 2.08 0.74 0.06
Switzerland 0.91 0.32 0.00 1.59 0.64 0.93
U.K. 0.85 0.39 0.01 1.67 0.77 1.00
Table 1b
The Performance of Single Equation Models(without W2).
Australia 0.29 1.00 0.52 0.42 NaN NaN NaN NaN 0.00 0.99 0.46 0.01
Belgium 0.89 1.01 0.55 0.74 0.94 1.01 0.57 0.88 NaN NaN NaN NaN
Canada 0.06 0.99 0.46 0.03 0.00 0.99 0.44 0.00 1.00 1.01 0.54 0.60
Denmark 0.79 1.00 0.54 0.73 0.11 1.00 0.53 0.58 0.24 1.02 0.55 0.33
Finland 0.11 1.03 0.63 0.98 0.13 1.01 0.57 0.96 0.80 1.02 0.58 0.91
France 0.90 1.01 0.56 0.92 0.38 1.04 0.65 0.95 0.17 1.00 0.49 0.11
Germany 0.74 1.01 0.55 0.78 0.67 1.01 0.55 0.76 0.11 1.00 0.50 0.03
Greece NaN NaN NaN NaN 0.62 0.99 0.47 0.01 0.07 1.01 0.53 0.00
Ireland 0.17 1.03 0.54 0.82 0.01 1.02 0.56 0.74 NaN NaN NaN NaN
Italy 0.60 1.00 0.53 0.55 0.01 1.01 0.54 0.42 0.03 1.04 0.55 0.21
Japan 0.48 1.01 0.54 0.52 0.76 1.03 0.54 0.36 0.04 1.02 0.56 0.84
New Zealand 0.65 1.00 0.51 0.53 NaN NaN NaN NaN 0.00 0.98 0.46 0.02
Portugal 0.72 1.00 0.50 0.24 0.00 1.01 0.54 0.48 NaN NaN NaN NaN
Spain 0.16 1.01 0.57 0.94 0.43 1.01 0.55 0.73 NaN NaN NaN NaN
Sweden 0.51 1.01 0.60 0.99 0.13 1.01 0.52 0.30 1.00 1.06 0.61 0.89
Switzerland 0.53 0.99 0.45 0.03 0.58 1.01 0.52 0.32 0.87 1.02 0.57 0.80
U.K. 0.38 1.00 0.54 0.76 0.29 1.01 0.56 0.81 0.63 1.01 0.57 0.80
Canada 0.16 1.00 0.51 0.29 0.39 1.00 0.51 0.22 1.00 1.01 0.54 0.60
Denmark 0.43 1.00 0.55 0.86 0.32 1.01 0.52 0.26 0.24 1.02 0.55 0.33
Finland 0.40 1.01 0.59 0.87 0.35 1.02 0.60 0.84 0.80 1.02 0.58 0.91
France 0.90 1.03 0.60 0.73 0.88 1.03 0.55 0.39 0.17 1.00 0.49 0.11
Germany 0.76 1.02 0.60 0.95 0.03 1.00 0.49 0.10 0.11 1.00 0.50 0.03
Italy 0.72 1.03 0.57 0.29 0.26 1.01 0.53 0.32 0.03 1.04 0.55 0.21
Japan 0.47 1.00 0.51 0.33 0.95 1.02 0.58 0.62 0.04 1.02 0.56 0.84
Sweden 0.65 1.01 0.56 0.89 0.48 1.01 0.58 0.83 1.00 1.06 0.61 0.89
Switzerland 0.76 1.00 0.48 0.19 0.49 1.00 0.48 0.10 0.87 1.02 0.57 0.80
U.K. 0.61 1.00 0.49 0.14 0.03 1.01 0.54 0.07 0.63 1.01 0.57 0.80
73
Table 2b

Canada 0.66 1.01 0.55 0.87 0.61 1.02 0.45 0.03
Denmark 0.03 0.99 0.43 0.04 0.50 1.01 0.53 0.35
Finland 0.07 1.00 0.49 0.20 0.11 1.00 0.55 0.22
France 0.03 1.01 0.54 0.37 0.23 1.05 0.55 0.07
Germany 0.16 1.00 0.51 0.29 0.00 1.04 0.50 0.05
Greece 0.50 1.01 0.56 0.79 0.00 0.92 0.46 0.14
Italy 0.24 1.02 0.56 0.67 0.00 1.02 0.45 0.06
Japan 0.12 1.01 0.53 0.51 0.00 0.98 0.32 0.03
Sweden 0.06 1.00 0.51 0.39 0.06 0.99 0.48 0.03
Switzerland 0.76 1.02 0.63 0.99 0.18 1.15 0.55 0.63
U.K. 0.03 1.00 0.54 0.54 0.10 1.00 0.49 0.08

Canada 0.12 1.01 0.54 0.58 0.61 1.02 0.45 0.03
Denmark 0.94 1.01 0.57 0.60 0.50 1.01 0.53 0.35
Finland 0.36 1.05 0.59 0.10 0.11 1.00 0.55 0.22
Germany 0.80 1.01 0.58 0.87 0.00 1.04 0.50 0.05
Greece 0.86 1.00 0.50 0.16 0.00 0.92 0.46 0.14
Italy 0.95 1.02 0.53 0.14 0.00 1.02 0.45 0.06
Sweden 0.16 1.04 0.60 0.76 0.06 0.99 0.48 0.03
Switzerland 0.44 1.00 0.52 0.30 0.18 1.15 0.55 0.63
U.K. 0.31 1.02 0.54 0.47 0.10 1.00 0.49 0.08
Table 3b
The performance of multiple equations models(without W2).
Australia 1.00 0.49 0.27 1.01 0.52 0.21

Canada 1.00 0.52 0.56 1.01 0.53 0.31
Denmark 1.00 0.52 0.43 1.01 0.58 0.89
Finland 0.99 0.44 0.06 1.01 0.54 0.53
France 0.99 0.44 0.02 1.00 0.49 0.04
Germany 1.00 0.52 0.49 1.00 0.52 0.26
Italy 1.01 0.52 0.43 1.02 0.55 0.22
Japan 1.00 0.50 0.27 1.01 0.54 0.66
New Zealand 1.00 0.50 0.26 1.01 0.52 0.09
Sweden 1.02 0.56 0.76 1.01 0.60 0.94
Switzerland 1.00 0.49 0.22 1.03 0.62 0.95
U.K. 1.00 0.53 0.66 1.01 0.55 0.73
74
Table 3b (continued )

France 1.00 0.51 0.40 1.95 0.99 1.00
Germany 1.00 0.48 0.12 1.94 0.72 1.00
Greece 1.00 0.50 0.27 0.81 0.15 0.00
Italy 0.99 0.46 0.06 1.37 0.82 0.00
Japan 0.99 0.44 0.04 1.46 0.63 0.97
New Zealand 1.01 0.54 0.50 1.75 0.74 0.91
Sweden 1.02 0.55 0.59 2.18 0.71 0.04
Switzerland 0.99 0.44 0.03 1.50 0.62 0.64
U.K. 1.00 0.50 0.24 1.67 0.78 1.00
Table 1c
Canada 0.56 1.00 0.49 0.16 0.01 1.00 0.47 0.03 1.00 1.01 0.54 0.67
Denmark 0.32 1.01 0.57 0.94 0.26 1.01 0.53 0.67 0.32 1.02 0.57 0.47
Finland 0.49 1.01 0.55 0.64 0.20 1.01 0.58 0.98 0.81 1.02 0.58 0.93
France 0.94 1.02 0.55 0.90 0.56 1.04 0.64 0.95 0.43 1.01 0.52 0.28
Germany 0.67 1.01 0.55 0.81 0.88 1.01 0.58 0.95 0.36 1.01 0.52 0.27
Italy 0.89 1.00 0.54 0.65 0.03 1.01 0.54 0.52 0.12 1.03 0.55 0.27
Japan 0.42 1.01 0.55 0.59 0.61 1.03 0.54 0.32 0.04 1.01 0.54 0.69
Sweden 0.65 1.01 0.56 0.93 0.23 1.00 0.52 0.27 1.00 1.05 0.61 0.96
Switzerland 0.38 1.00 0.48 0.16 0.58 1.01 0.52 0.25 0.77 1.01 0.57 0.80
U.K. 0.12 1.01 0.55 0.79 0.49 1.01 0.56 0.84 0.34 1.01 0.58 0.87
Canada 0.10 1.00 0.50 0.29 0.52 1.00 0.52 0.36 1.00 1.01 0.54 0.67
Denmark 0.06 1.00 0.56 0.91 0.51 1.01 0.54 0.59 0.32 1.02 0.57 0.47
Finland 0.53 1.02 0.57 0.83 0.72 1.02 0.65 0.99 0.81 1.02 0.58 0.93
France 0.29 1.01 0.55 0.18 0.82 1.04 0.60 0.60 0.43 1.01 0.52 0.28
Germany 0.71 1.02 0.58 0.86 0.07 1.00 0.51 0.25 0.36 1.01 0.52 0.27
Italy 0.59 1.03 0.57 0.24 0.32 1.01 0.54 0.45 0.12 1.03 0.55 0.27
Japan 0.76 1.00 0.49 0.23 0.70 1.01 0.54 0.40 0.04 1.01 0.54 0.69
Sweden 0.00 1.00 0.56 0.87 0.78 1.01 0.58 0.91 1.00 1.05 0.61 0.96
Switzerland 0.62 1.00 0.50 0.37 0.49 1.00 0.48 0.12 0.77 1.01 0.57 0.80
U.K. 0.91 1.00 0.51 0.31 0.12 1.01 0.53 0.22 0.34 1.01 0.58 0.87
75
Table 2c

Canada 0.66 1.01 0.55 0.87 0.79 1.00 0.42 0.02
Denmark 0.03 0.99 0.43 0.04 0.71 1.01 0.53 0.38
Finland 0.07 1.00 0.49 0.20 0.17 1.00 0.48 0.24
France 0.03 1.01 0.54 0.37 0.15 1.01 0.48 0.14
Germany 0.16 1.00 0.51 0.29 0.34 1.00 0.43 0.11
Greece 0.50 1.01 0.56 0.79 0.00 0.89 0.44 0.04
Italy 0.24 1.02 0.56 0.67 0.00 0.96 0.45 0.02
Japan 0.12 1.01 0.53 0.51 0.00 0.91 0.31 0.00
Sweden 0.06 1.00 0.51 0.39 0.12 0.99 0.46 0.03
Switzerland 0.76 1.02 0.63 0.99 0.38 1.06 0.47 0.24
U.K. 0.03 1.00 0.54 0.54 0.12 1.01 0.50 0.52

Canada 0.12 1.01 0.54 0.58 0.79 1.00 0.42 0.02
Denmark 0.94 1.01 0.57 0.60 0.71 1.01 0.53 0.38
Finland 0.36 1.05 0.59 0.10 0.17 1.00 0.48 0.24
Germany 0.80 1.01 0.58 0.87 0.34 1.00 0.43 0.11
Greece 0.86 1.00 0.50 0.16 0.00 0.89 0.44 0.04
Italy 0.95 1.02 0.53 0.14 0.00 0.96 0.45 0.02
Sweden 0.16 1.04 0.60 0.76 0.12 0.99 0.46 0.03
Switzerland 0.44 1.00 0.52 0.30 0.38 1.06 0.47 0.24
U.K. 0.31 1.02 0.54 0.47 0.12 1.01 0.50 0.52
Table 3c
The Performance of Multiple Equations Models (without W3).
Australia 1.00 0.53 0.71 1.01 0.52 0.30

Canada 1.00 0.51 0.39 1.01 0.54 0.49
Denmark 1.00 0.52 0.41 1.00 0.53 0.59
Finland 0.99 0.46 0.13 1.01 0.54 0.57
France 0.99 0.41 0.01 0.99 0.49 0.02
Germany 1.00 0.51 0.46 1.01 0.53 0.37
Italy 1.01 0.54 0.61 1.02 0.53 0.17
Japan 1.00 0.47 0.05 1.01 0.55 0.74
New Zealand 1.00 0.48 0.18 1.01 0.54 0.23
Sweden 1.01 0.55 0.75 1.01 0.57 0.82
Switzerland 1.00 0.48 0.17 1.03 0.60 0.92
U.K. 1.00 0.51 0.38 1.01 0.56 0.76
76
Table 3c (continued )

France 1.00 0.54 0.81 1.89 0.99 1.00
Germany 1.00 0.47 0.06 1.97 0.72 1.00
Greece 1.00 0.48 0.16 0.75 0.02 0.00
Italy 1.00 0.46 0.09 1.41 0.84 0.00
Japan 1.00 0.47 0.12 1.47 0.63 0.95
New Zealand 1.00 0.55 0.68 1.59 0.76 0.77
Sweden 1.01 0.55 0.62 1.76 0.69 0.08
Switzerland 0.99 0.43 0.03 1.53 0.61 0.42
U.K. 1.01 0.55 0.68 1.67 0.78 1.00
Table 1d
Canada 0.33 1.00 0.48 0.14 0.01 1.00 0.49 0.10 0.90 1.02 0.55 0.83
Denmark 0.33 1.01 0.57 0.95 0.38 1.01 0.54 0.75 0.29 1.02 0.56 0.38
Finland 0.94 1.01 0.57 0.79 0.21 1.01 0.56 0.94 0.90 1.02 0.57 0.91
France 0.80 1.01 0.55 0.91 0.41 1.02 0.58 0.76 0.84 1.03 0.56 0.45
Germany 0.51 1.01 0.55 0.79 0.94 1.01 0.58 0.97 0.80 1.01 0.56 0.84
Italy 0.72 1.00 0.53 0.60 0.01 1.01 0.55 0.70 0.07 1.02 0.52 0.17
Japan 0.36 1.01 0.54 0.62 0.59 1.02 0.54 0.42 0.16 1.01 0.56 0.91
Sweden 0.67 1.01 0.55 0.85 0.32 1.00 0.51 0.24 1.00 1.04 0.63 0.99
Switzerland 0.56 1.00 0.49 0.25 0.44 1.01 0.52 0.20 0.76 1.01 0.55 0.69
U.K. 0.12 1.01 0.55 0.75 0.76 1.01 0.56 0.90 0.15 1.02 0.58 0.87
Canada 0.20 1.00 0.50 0.31 0.59 1.00 0.52 0.44 0.90 1.02 0.55 0.83
Denmark 0.05 1.00 0.55 0.85 0.61 1.01 0.55 0.78 0.29 1.02 0.56 0.38
Finland 0.49 1.01 0.58 0.82 0.92 1.03 0.65 1.00 0.90 1.02 0.57 0.91
France 0.48 1.02 0.57 0.24 0.82 1.04 0.62 0.70 0.84 1.03 0.56 0.45
Germany 0.82 1.02 0.60 0.94 0.16 1.01 0.53 0.39 0.80 1.01 0.56 0.84
Italy 0.95 1.03 0.58 0.29 0.58 1.01 0.54 0.47 0.07 1.02 0.52 0.17
Japan 0.77 1.00 0.52 0.51 0.37 1.02 0.54 0.53 0.16 1.01 0.56 0.91
Sweden 0.11 1.00 0.56 0.92 0.94 1.01 0.58 0.93 1.00 1.04 0.63 0.99
Switzerland 0.51 1.00 0.50 0.30 0.45 1.00 0.48 0.12 0.76 1.01 0.55 0.69
U.K. 0.79 1.00 0.52 0.35 0.23 1.01 0.56 0.59 0.15 1.02 0.58 0.87
77
Table 2d

Canada 0.66 1.01 0.55 0.87 0.72 0.99 0.41 0.06
Denmark 0.03 0.99 0.43 0.04 0.84 1.00 0.52 0.16
Finland 0.07 1.00 0.49 0.20 0.46 1.00 0.53 0.35
France 0.03 1.01 0.54 0.37 0.65 1.00 0.47 0.15
Germany 0.16 1.00 0.51 0.29 0.12 0.99 0.41 0.03
Greece 0.50 1.01 0.56 0.79 0.00 0.92 0.43 0.01
Italy 0.24 1.02 0.56 0.67 0.00 0.98 0.47 0.00
Japan 0.12 1.01 0.53 0.51 0.07 0.93 0.41 0.03
Sweden 0.06 1.00 0.51 0.39 0.50 0.99 0.47 0.07
Switzerland 0.76 1.02 0.63 0.99 0.15 1.04 0.42 0.20
U.K. 0.03 1.00 0.54 0.54 1.00 1.01 0.52 0.28

Canada 0.12 1.01 0.54 0.58 0.72 0.99 0.41 0.06
Denmark 0.94 1.01 0.57 0.60 0.84 1.00 0.52 0.16
Finland 0.36 1.05 0.59 0.10 0.46 1.00 0.53 0.35
Germany 0.80 1.01 0.58 0.87 0.12 0.99 0.41 0.03
Greece 0.86 1.00 0.50 0.16 0.00 0.92 0.43 0.01
Italy 0.95 1.02 0.53 0.14 0.00 0.98 0.47 0.00
Sweden 0.16 1.04 0.60 0.76 0.50 0.99 0.47 0.07
Switzerland 0.44 1.00 0.52 0.30 0.15 1.04 0.42 0.20
U.K. 0.31 1.02 0.54 0.47 1.00 1.01 0.52 0.28
Table 3d
The performance of multiple equations models (without W4).
Australia 1.00 0.55 0.88 1.01 0.52 0.34

Canada 1.00 0.51 0.53 1.01 0.54 0.56
Denmark 1.00 0.53 0.56 1.00 0.52 0.54
Finland 1.00 0.50 0.38 1.01 0.54 0.62
France 1.00 0.50 0.36 1.00 0.50 0.02
Germany 1.00 0.56 0.90 1.01 0.54 0.49
Italy 1.01 0.57 0.83 1.02 0.54 0.15
Japan 1.00 0.54 0.80 1.00 0.53 0.63
New Zealand 1.01 0.57 0.89 1.02 0.56 0.55
Sweden 1.01 0.55 0.75 1.00 0.54 0.58
Switzerland 1.00 0.53 0.55 1.03 0.60 0.91
U.K. 1.00 0.53 0.62 1.01 0.55 0.80
78
Table 3d (continued )

France 1.00 0.56 0.94 1.92 1.00 1.00
Germany 1.00 0.48 0.13 1.97 0.72 1.00
Greece 1.00 0.48 0.15 0.68 0.00 0.00
Italy 1.00 0.50 0.31 1.42 0.83 0.00
Japan 1.00 0.50 0.27 1.49 0.63 0.91
New Zealand 1.00 0.52 0.44 1.68 0.87 0.95
Sweden 1.00 0.53 0.51 1.70 0.79 0.39
Switzerland 0.99 0.43 0.03 1.61 0.67 1.00
U.K. 1.00 0.53 0.54 1.65 0.80 1.00
Table 1e
Canada 0.33 1.00 0.48 0.15 0.00 0.99 0.47 0.03 0.85 1.01 0.56 0.85
Denmark 0.52 1.01 0.56 0.92 0.12 1.00 0.52 0.46 0.21 1.02 0.55 0.30
Finland 0.88 1.01 0.57 0.80 0.30 1.01 0.56 0.91 0.69 1.02 0.56 0.86
France 0.68 1.01 0.55 0.90 0.48 1.02 0.58 0.79 1.00 1.02 0.57 0.68
Germany 0.44 1.01 0.54 0.68 0.85 1.01 0.55 0.85 0.81 1.01 0.56 0.89
Italy 0.57 1.00 0.53 0.69 0.03 1.01 0.55 0.75 0.11 1.01 0.52 0.21
Japan 0.30 1.01 0.54 0.60 0.71 1.03 0.55 0.55 0.45 1.01 0.58 0.98
Sweden 0.67 1.01 0.55 0.86 0.32 1.00 0.51 0.24 0.69 1.04 0.62 0.92
Switzerland 0.46 1.00 0.49 0.23 0.12 1.00 0.50 0.10 0.83 1.01 0.55 0.67
U.K. 0.23 1.01 0.55 0.81 0.36 1.01 0.58 0.93 0.26 1.02 0.59 0.85
Canada 0.18 1.00 0.50 0.29 0.94 1.00 0.53 0.55 0.85 1.01 0.56 0.85
Denmark 0.06 1.00 0.54 0.82 0.46 1.01 0.54 0.74 0.21 1.02 0.55 0.30
Finland 0.34 1.01 0.56 0.69 0.42 1.02 0.64 0.99 0.69 1.02 0.56 0.86
France 0.53 1.02 0.58 0.26 0.95 1.03 0.59 0.61 1.00 1.02 0.57 0.68
Germany 0.88 1.02 0.59 0.91 0.28 1.02 0.55 0.64 0.81 1.01 0.56 0.89
Italy 0.95 1.04 0.57 0.37 0.72 1.02 0.55 0.61 0.11 1.01 0.52 0.21
Japan 0.53 1.00 0.54 0.68 0.60 1.02 0.56 0.61 0.45 1.01 0.58 0.98
Sweden 0.23 1.00 0.57 0.96 0.80 1.01 0.57 0.92 0.69 1.04 0.62 0.92
Switzerland 0.55 1.00 0.51 0.40 0.46 1.00 0.49 0.17 0.83 1.01 0.55 0.67
U.K. 0.74 1.00 0.52 0.40 0.04 1.01 0.53 0.18 0.26 1.02 0.59 0.85
79
Table 2e

Canada 0.66 1.01 0.55 0.87 0.37 0.99 0.43 0.01
Denmark 0.03 0.99 0.43 0.04 0.91 1.00 0.51 0.16
Finland 0.07 1.00 0.49 0.20 0.67 1.00 0.52 0.19
France 0.03 1.01 0.54 0.37 1.00 1.00 0.43 0.13
Germany 0.16 1.00 0.51 0.29 0.23 0.99 0.44 0.01
Greece 0.50 1.01 0.56 0.79 0.00 0.95 0.44 0.01
Italy 0.24 1.02 0.56 0.67 0.01 0.99 0.49 0.01
Japan 0.12 1.01 0.53 0.51 0.61 0.92 0.43 0.06
Sweden 0.06 1.00 0.51 0.39 0.83 0.99 0.48 0.05
Switzerland 0.76 1.02 0.63 0.99 0.54 1.02 0.40 0.16
U.K. 0.03 1.00 0.54 0.54 0.38 1.00 0.54 0.05

Canada 0.12 1.01 0.54 0.58 0.37 0.99 0.43 0.01
Denmark 0.94 1.01 0.57 0.60 0.91 1.00 0.51 0.16
Finland 0.36 1.05 0.59 0.10 0.67 1.00 0.52 0.19
Germany 0.80 1.01 0.58 0.87 0.23 0.99 0.44 0.01
Greece 0.86 1.00 0.50 0.16 0.00 0.95 0.44 0.01
Italy 0.95 1.02 0.53 0.14 0.01 0.99 0.49 0.01
Sweden 0.16 1.04 0.60 0.76 0.83 0.99 0.48 0.05
Switzerland 0.44 1.00 0.52 0.30 0.54 1.02 0.40 0.16
U.K. 0.31 1.02 0.54 0.47 0.38 1.00 0.54 0.05
Table 3e
The performance of multiple equations models (without W5).
Australia 1.00 0.55 0.89 1.01 0.52 0.34

Canada 1.00 0.55 0.86 1.01 0.54 0.60
Denmark 1.01 0.56 0.84 1.00 0.49 0.34
Finland 1.00 0.53 0.61 1.01 0.55 0.72
France 1.00 0.53 0.62 1.01 0.52 0.05
Germany 1.00 0.55 0.86 1.01 0.55 0.62
Italy 1.01 0.57 0.81 1.02 0.54 0.26
Japan 1.00 0.54 0.81 1.00 0.53 0.58
New Zealand 1.01 0.57 0.88 1.03 0.58 0.78
Sweden 1.01 0.55 0.74 1.00 0.54 0.58
Switzerland 1.00 0.52 0.49 1.03 0.59 0.90
U.K. 1.00 0.54 0.75 1.01 0.56 0.87
80
Table 3e (continued )

France 1.00 0.53 0.68 1.97 1.00 1.00
Germany 1.00 0.52 0.60 1.91 0.72 1.00
Greece 1.00 0.52 0.51 0.86 0.20 0.00
Italy 1.00 0.54 0.59 1.13 0.67 0.00
Japan 1.00 0.51 0.42 1.41 0.62 0.71
New Zealand 1.00 0.52 0.46 1.71 0.88 0.99
Sweden 1.01 0.57 0.65 1.51 0.79 0.98
Switzerland 1.00 0.46 0.10 1.66 0.79 1.00
U.K. 1.00 0.47 0.12 1.61 0.81 1.00
References
Alquist, R., & Chinn, M. D. (2008). Conventional and unconventional approaches to exchange rate modelling and assessment. International Journal of Finance & Eco-
nomics, 13(1), 2–13.
Aussem, A., Campbell, J., & Murtagh, F. (1998). Wavelet-based feature extraction and decomposition strategies for financial forecasting. Journal of Compuational
Intelligence in Finance, 6, 5–12.
Caraiani, P. (2015). Estimating dsge models across time and frequency. Journal of Macroeconomics, 44(C), 33–49.
Caraiani, P. (2016). Money and output causality: A structural approach. International Review of Economics & Finance, 42(C), 220–236.
Cerra, V., & Saxena, S. C. (2010). The monetary model strikes back: Evidence from the world. Journal of International Economics, 81(2), 184–196.
Cheung, Y.-W., Chinn, M. D., & Pascual, A. G. (2005). Empirical exchange rate models of the nineties: Are any fit to survive? Journal of International Money and Finance,
24(7), 1150–1175.
Chinn, M. D., & Meese, R. A. (1995). Banking on currency forecasts: How predictable is change in money? Journal of International Economics, 38(1–2), 161–178.
Clark, T. E., & West, K. D. (2006). Using out-of-sample mean squared prediction errors to test the martingale difference hypothesis. Journal of Econometrics, 135(1–2),
155–186.
Engel, C., & West, K. D. (2005). Exchange rates and fundamentals. Journal of Political Economy, 113(3), 485–517.
Engel, C., & West, K. D. (2006). Taylor rules and the Deutschmark: Dollar real exchange rate. Journal of Money, Credit and Banking, 38(5), 1175–1194.
Ferraro, D., Rogoff, K., & Rossi, B. (2011). Can oil prices forecast exchange rates?. Working Papers 11-05. Duke University, Department of Economics. URL https://ideas.
repec.org/p/duk/dukeec/11-05.html.
Fisher, I. (1896). Appreciation and interest. New York: Macmillan for the American Economic Association.
Frenkel, J. A. (1976). A monetary approach to the exchange rate: Doctrinal aspects and empirical evidence. Scandinavian Journal of Economics, 78(2), 200–224.
Gallegati, M., Gallegati, M., Ramsey, J. B., & Semmler, W. (2011). The US wage Phillips curve across frequencies and over time. Oxford Bulletin of Economics and
Statistics, 73(4), 489–508.
Gencay, R., Selcuk, F., & Whitcher, B. J. (2001). An introduction to wavelets and other filtering methods in finance and economics. Cambridge: Cambridge University Press.
Mark, N. C. (1995). Exchange rates and fundamentals: Evidence on long-horizon predictability. American Economic Review, 85(1), 201–218.
Mark, N. C., & Sul, D. (2001). Nominal exchange rates and monetary fundamentals: Evidence from a small post-Bretton woods panel. Journal of International Economics,
53(1), 29–52.
Meese, R. A., & Rogoff, K. (1983). Empirical exchange rate models of the seventies : Do they fit out of sample? Journal of International Economics, 14(1–2), 3–24.
Michis, A. A. (2014). Time scale evaluation of economic forecasts. Economics Letters, 123(3), 279–281.
Molodtsova, T., & Papell, D. H. (2009). Out-of-sample exchange rate predictability with Taylor rule fundamentals. Journal of International Economics, 77(2), 167–180.
Molodtsova, T., & Papell, D. H. (2013). Taylor rule exchange rate forecasting during the financial crisis. NBER International Seminar on Macroeconomics, 9(1), 55–97.
Mussa, M. (1976). The exchange rate, the balance of payments and monetary and fiscal policy under a regime of controlled floating. Scandinavian Journal of Economics,
78(2), 229–248.
Percival, D., & Walden, A. T. (2000). Wavelets methods for time series analysis. London: Academic Press.
Rabanal, P., & Rubio-Ramrez, J. F. (2015). Can international macroeconomic models explain low-frequency movements of real exchange rates? Journal of International
Economics, 96(1), 199–211.
Rapach, D. E., & Wohar, M. E. (2006). The out-of-sample forecasting performance of nonlinear models of real exchange rate behavior. International Journal of Forecasting,
22(2), 341–361.
Renaud, O., Starck, J. L., & Murtagh, F. (2005). Wavelet-based combined signal filtering and prediction. IEEE Transactions on Systems, Man, and Cybernetics, B
(Cybernetics), 35(6), 1241–1251.
Rossi, B. (2005a). Confidence intervals for half-life deviations from purchasing power parity. Journal of Business & Economic Statistics, 23, 432–442.
Rossi, B. (2005c). Testing long-horizon predictive ability with high persistence, and the Meese-Rogoff puzzle. International Economic Review, 46(1), 61–92.
Rossi, B. (2013). Exchange rate predictability. Journal of Economic Literature, 51(4), 1063–1119.
Rossi, B. (2005b October). Optimal tests for nested model selection with underlying parameter instability. Econometric Theory, 21(05), 962–990.
Rua, A. (2011). A wavelet approach for factoraugmented forecasting. Journal of Forecasting, 30(7), 666–678.
Soltani, S., Boichu, D., Simard, P., & Canu, S. (2000). The long-term memory prediction by multiscale decomposition. Signal Processing, 80, 2195–2205.
Taylor, J. B. (1993). Discretion versus policy rules in practice. Carnegie-Rochester Conference Series on Public Policy, 39(1), 195–214.
Wright, J. H. (2008). Bayesian model averaging and exchange rate forecasts. Journal of Econometrics, 146(2), 329–341.
Zheng, G., Starck, J.-L., Campbell, J., & Murtagh, F. (1999). The wavelet transform for filtering financial data streams. Journal of Compuational Intelligence in Finance, 7,
18–35.
81

International Review of Economics and Finance: Petre Caraiani

Enviado por

Dados do documento

Título original

Direitos autorais

Formatos disponíveis

Compartilhar este documento

Compartilhar ou incorporar documento

Opções de compartilhamento

Você considera este documento útil?

Este conteúdo é inapropriado?

Direitos autorais:

Formatos disponíveis

International Review of Economics and Finance: Petre Caraiani

Enviado por

Direitos autorais:

Formatos disponíveis

International Review of Economics and Finance 51 (2017) 60–81

Contents lists available at ScienceDirect

International Review of Economics and Finance

Evaluating exchange rate forecasts along time and frequency

E-mail address: caraiani@ipe.ro.

mt pt ¼ ηitþ1 þ ϕyt (3)

where π t is the inﬂation rate while ygap

4.1. Maximal overlap discrete wavelet transform

and unit energy:

for all nonzero integers n.

for t ¼ 0; …; N=2j 1, for the wavelet coefﬁcients, and

vJ;t ¼ 2J=2 ~vJ;2J ðjþ1Þ1 (15)

~j;t is the component t of d

for j ¼ 1; ; J. The smooth component can be obtained by using:

4.2. The redundant Haar wavelet transform

c0 ðkÞ ¼ < f ðxÞ; ϕðx kÞ > (19)

wj ðkÞ ¼ cj1 ðkÞ cj ðkÞ (22)

But this can be alternatively written as:

4.3. Forecast evaluation of the different frequencies

5.1. Model performances

5.2. Further statistics

5.3. A comparison with the results in the literature

Appendix A. Data series samples across countries

Appendix B. In and out-of-sample performance across details

Fig. B.1. Scatterplot GC: The Series without W1.

Fig. B.2. Scatterplot: The Series without W1.

Fig. B.3. Scatterplot GC: The Series without W2.

Fig. B.4. Scatterplot: The Series without W2.

Fig. B.5. Scatterplot GC: The Series without W3.

Fig. B.6. Scatterplot: The Series without W3.

Fig. B.7. Scatterplot GC: The Series without W4.

Fig. B.8. Scatterplot: The Series without W4.

Fig. B.9. Scatterplot GC: The Series without W5.

Fig. B.10. Scatterplot: The Series without W5.

Appendix C. The performance of the models

Country 1.UIRP model 2.PPP Mode 3.Monetary Model

h ¼ 1 month h ¼ 1 month h ¼ 1 month

GC RMSER pvDM pvCW GC RMSER pvDM pvCW GC RMSER pvDM pvCW

h ¼ 4 years h ¼ 4 years h ¼ 4 years

Country 1.Monetary ECM model 2.Taylor Model

GC RMSER pvDM pvCW GC RMSER pvDM pvCW

Australia 0.31 1.02 0.60 0.99 NaN NaN NaN NaN

Australia 0.47 1.00 0.54 0.58 NaN NaN NaN NaN

Country 1.Monetary Panel model 2.BMA Model

RMSER pvDM pvCW RMSER pvDM pvCW

Australia 0.91 0.33 0.00 1.00 0.51 0.15

Country 1.Monetary Panel model 2.BMA Model

RMSER pvDM pvCW RMSER pvDM pvCW

Australia NaN NaN NaN 1.45 0.77 1.00

Country 1.UIRP model 2.PPP Mode 3.Monetary Model

h ¼ 1 month h ¼ 1 month h ¼ 1 month

GC RMSER pvDM pvCW GC RMSER pvDM pvCW GC RMSER pvDM pvCW

h ¼ 4 years h ¼ 4 years h ¼ 4 years

Country 1.Monetary ECM model 2.Taylor Model

GC RMSER pvDM pvCW GC RMSER pvDM pvCW

Australia 0.31 1.02 0.60 0.99 NaN NaN NaN NaN

Australia 0.47 1.00 0.54 0.58 NaN NaN NaN NaN

Country 1.Monetary Panel model 2.BMA Model

RMSER pvDM pvCW RMSER pvDM pvCW

Australia 1.00 0.49 0.27 1.01 0.52 0.21

Country 1.Monetary Panel model 2.BMA Model

RMSER pvDM pvCW RMSER pvDM pvCW

Australia NaN NaN NaN 1.45 0.77 1.00