WORKING PAPER

Using low frequency information for NORGES BANK

RESEARCH
predicting high frequency variables
13 | 2015

AUTHORS:

CLAUDIA FORONI,
PIERRE GUÉRIN AND
MASSIMILIANO
MARCELLINO
Working papers fra Norges Bank, fra 1992/1 til 2009/2 kan bestilles over e-post: NORGES BANK
[email protected]
WORKING PAPER
XX | 2014
Fra 1999 og senere er publikasjonene tilgjengelige på www.norges-bank.no
RAPPORTNAVN
Working papers inneholder forskningsarbeider og utredninger som vanligvis ikke har fått
sin endelige form. Hensikten er blant annet at forfatteren kan motta kommentarer fra
kolleger og andre interesserte. Synspunkter og konklusjoner i arbeidene står for
forfatternes regning.

Working papers from Norges Bank, from 1992/1 to 2009/2 can be ordered by e-mail:

[email protected]

Working papers from 1999 onwards are available on www.norges-bank.no

Norges Bank’s working papers present research projects and reports (not usually in their
final form) and are intended inter alia to enable the author to benefit from the comments of
colleagues and other interested parties. Views and conclusions expressed in working
papers are the responsibility of the authors alone.

ISSN 1502-8143 (online)

ISBN 978-82-7553-878-7 (online)

2
 Using Low Frequency Information for Predicting High
Frequency Variables ∗
†
Claudia Foroni Pierre Guérin‡ Massimiliano Marcellino§

October 29, 2015

Abstract
We analyze how to incorporate low frequency information in models for predicting
high frequency variables. In doing so, we introduce a new model, the reverse unre-
stricted MIDAS (RU-MIDAS), which has a periodic structure but can be estimated
by simple least squares methods and used to produce forecasts of high frequency
variables that also incorporate low frequency information. We compare this model
with two versions of the mixed frequency VAR, which so far had been only applied
to study the reverse problem, that is, using the high frequency information for pre-
dicting low frequency variables. We then implement a simulation study to evaluate
the relative forecasting ability of the alternative models in finite samples. Finally,
we conduct several empirical applications to assess the relevance of quarterly survey
data for forecasting a set of monthly macroeconomic indicators. Overall, it turns out
that low frequency information is important, particularly so when it is just released.

Keywords: Mixed-Frequency VAR models, temporal aggregation, MIDAS models.

JEL Classification Code: E37, C53.

∗
This Working Paper should not be reported as representing the views of Norges Bank or Bank of
Canada. The views expressed are those of the authors and do not necessarily reflect those of Norges Bank
or Bank of Canada. We would like to thank Marta Banbura, Chiara Osbat, seminar participants at the
ECB, the 23rd SNDE symposium, the 2nd IAAE annual conference, the Barcelona GSE Summer Forum
on Time Series Analysis in Macro and Finance, and the EABCN conference on ”Econometric methods for
business cycle analysis, forecasting and policy simulations” for helpful comments.
†
Norges Bank
‡
Bank of Canada
§
Bocconi University, IGIER and CEPR

1
1 Introduction

A large and increasing literature is dealing with models that explicitly account for data of
different frequencies, especially in a forecasting context. Different classes of models have
been developed to tackle the different sampling frequencies at which macroeconomic and
financial indicators are available. First, a common choice is to cast the model in state
space form and use the Kalman filter for estimation and forecasting (see, among many
others, Mariano and Murasawa (2003), Giannone et al. (2008), Aruoba et al. (2009)).
Alternatively, in a Bayesian context, the estimation of models with mixed-frequency data
has been studied, for example, by Eraker et al. (2015) and Schorfheide and Song (2015).
Second, Ghysels (2014) introduced a different class of mixed-frequency VAR model, in
which the vector of dependent variables includes both high-frequency and low-frequency
variables, with the former stacked according to the timing of the data release (see also
Blasques et al. (2014) for an application in the context of a small-scale factor model). The
model can be estimated by simple OLS but stacking increases the number of regressors.
Bayesian methods can be used to handle this issue when the number of degrees of freedom
is very small. Finally, in a univariate context, the MIDAS approach by Ghysels et al.
(2006) directly links low-frequency to high-frequency data. MIDAS models have been
extensively used in macroeconomic forecasting (see, e.g., Clements and Galvao (2008) and
Clements and Galvao (2009) for early contributions). MIDAS models, being non-linear,
require a form of NLS estimation, which increases substantially the computational costs in
specifications with more than one high frequency explanatory variable. The unrestricted
MIDAS, U-MIDAS, model of Foroni et al. (2013b) can instead be estimated by simple
OLS and therefore handle several high frequency explanatory variables. However, it works
particularly well only when the frequency mismatch is low, for example in the quarterly /
monthly case.

The attention in the literature so far has mostly been on how to exploit high-frequency
information to improve forecasts of low-frequency variables. One notable exception is
Dal Bianco et al. (2012), who analyze how macroeconomic fundamentals can improve the
forecasts of the euro-dollar exchange rate at the weekly frequency. In doing so, they estimate
a mixed-frequency VAR cast in state-space form à la Mariano and Murasawa (2010).

The aim of our paper is to study in a systematic way how to incorporate low-frequency
information in forecasting models for high-frequency variables. We propose a new model
where the high-frequency dependent variable is dynamically linked with the low-frequency

2
explanatory variable(s). We denote this model as Reverse Unrestricted MIDAS (RU-
MIDAS), since it is the counterpart of the U-MIDAS model of Foroni et al. (2013b). In
particular, we also discuss how RU-MIDAS regressions can be derived in a general lin-
ear dynamic framework, following similar steps to the analysis presented in Foroni et al.
(2013b) for the U-MIDAS regression. Next, this new model is analyzed in comparison
with two other standard mixed-frequency methods: a mixed-frequency VAR cast in state-
space form (see, e.g., Mariano and Murasawa (2010)), in which the low-frequency series
are treated as high-frequency variables with missing observations, and the mixed-frequency
VAR suggested by Ghysels (2014), where the high-frequency and low-frequency variables
are stacked together in the vector of dependent variables. Both methods are well known
but, as mentioned, they have been mostly applied to the analysis of low-frequency variables,
while here we focus on predicting high-frequency variables.

We then evaluate the forecasting ability of the different models in a Monte Carlo study,
focusing on the monthly / quarterly case that is particularly relevant for macroeconomic
applications. We consider different data generating processes (DGPs), based on bivariate
VARs either at high frequency or at mixed-frequency. We find that the forecasting per-
formance for the monthly variable of all the models we consider drastically improves right
after the release of the quarterly variable. In the same vein, the forecasting performance
deteriorates as we move away from the release of the quarterly variable and the monthly AR
benchmark becomes difficult to beat. Moreover, there is no clear ranking of the different
low-to-high frequency models, since their relative performance varies across horizons and
specifications.

Finally, to assess the performance of the alternative methods in empirical applications,

we evaluate the relevance of surveys available only at quarterly frequency to forecast the
corresponding monthly variables. It is well documented in the literature that expectations
contained in survey data help in improving the accuracy of the forecasts for key macroe-
conomic variables (see, for example, Chun (2011), Chernov and Mueller (2012) and Stark
(2010)). In particular, we use the Survey of Professional Forecasters (SPF), which is recog-
nized as a very accurate survey (see Stark (2010), who performs an extensive evaluation of
the SPF). Moreover, surveys of professional forecasters receive a lot of attention from both
market participants and policy makers as they are often seen as a good proxy of agents’ ex-
pectations and future economic outcomes. Ghysels and Wright (2009) use mixed-frequency
data models to obtain high-frequency estimates of survey releases, and find they provide
helpful estimates for future economic activity. In contrast, in this paper, we directly use the

3
information of the quarterly survey (that is, without conducting any prior disaggregation)
to improve the forecasts of important monthly U.S. variables: CPI inflation, industrial
production growth and real personal consumption expenditures.

As a second application, we assess whether the SPF has predictive power for exchange
rates. The literature on exchange rates is vast and generally agrees on the limited use of
macroeconomic fundamentals to predict exchange rates, a feature also known as the Meese
and Rogoff puzzle (see Meese and Rogoff (1983b), Engel and West (2005) and Rime et al.
(2010) among others). Surveys instead represent expected future fundamentals, which can
perform better in forecasting than the current macroeconomic fundamentals. In particular,
we evaluate whether expectations on the three-month T-Bill rate and CPI inflation have
predictive power for the trade-weighted average of the foreign exchange value of the U.S.
Dollar against major currencies. This choice of predictors is motivated by the purchasing
power parity and the uncovered interest rate parity relations (see Rossi (2013) for a review
of the literature on exchange rate forecasting).1

In our empirical applications, we obtain satisfactory results from the use of low frequency
(quarterly) survey information, particularly in forecasting monthly inflation and industrial
production, at almost every horizon. Moreover, we succeed in outperforming the random
walk benchmark in forecasting the monthly nominal U.S. trade-weighted exchange rate,
especially when surveys on inflation are employed. However, it remains difficult to find
a single preferred high frequency model, confirming the evidence from the Monte Carlo
experiments.

The paper is organized as follows. In Section 2, we introduce the RU-MIDAS model. In

Section 3, we present the alternative models we consider, i.e., the MF-VAR cast in state-
space form and the MF-VAR in stacked form. In Section 4, we discuss the Monte Carlo
experiments. In Section 5, we present the empirical applications that forecast monthly
economic indicators using the quarterly SPF. In Section 6, we summarize the main results
and conclude.
1
We focus on the trade-weighted exchange rate since it is a relevant indicator of the competitiveness of
a country. Moreover, for bilateral exchange rates, macroeconomic differentials (or expectations of macroe-
conomic differentials) would likely be more relevant than expectations related to the U.S. economy only.
However, survey-based expectations for other countries are not available over a long enough sample or not
fully comparable across countries, so that we refrained from considering the case of forecasting bilateral
exchange rates with survey data.

4
2 RU-MIDAS

In this section, we derive the Reverse Unrestricted MIDAS (RU-MIDAS) regression

approach from a general dynamic linear model, discuss its estimation, and consider the use
of RU-MIDAS as a forecasting device.

Let us denote the high frequency (HF) time index as t = 0, k1 , k2 , ..., k−1
k
, 1, ..., where k
denotes the frequency mismatch, and the low frequency (LF) time index as t = 0, 1, 2, ....
The variable x can be observed in HF (for each t) while the variable y ∗ can be only observed
in LF (every k periods). More generally, if we denote by L the HF lag operator so that
Lj xt = xt−j/k , then we can introduce the operator ω(L):

ω(L) = ω0 + ω1 L + ... + ωk−1 Lk−1 , (1)

which characterizes the temporal aggregation scheme applied to the y ∗ variable. For exam-
ple, ω(L) = 1 + L + ... + Lk−1 in the case of flow variables and ω(L) = 1 for stock variables.
What we observe in LF is yt = ω(L)yt∗ .

We assume that the variable x is generated by an AR(p) process with the variable y ∗
as an exogenous regressor:
c(L)xt = d(L)yt∗ + ext , (2)
where d(L) = d1 L + ... + dp Lp , c(L) = I − c1 L − ... − cp Lp , and the errors are white noise.
∗ ∗
For simplicity, we suppose that the starting values y−p/k , ..., y−1/k and x−p/k , ..., x−1/k are
2
all fixed and equal to zero.

Let us now introduce the LF lag operator, Z, with Z = Lk so that Z j yt = yt−j , and
define a polynomial in the HF lag operator, γ0 (L), such that the product g0 (L) = γ0 (L)d(L)
only contains powers of Lk = Z, so that g0 (L) = g0 (Lk ) = g0 (Z). Multiplying both sides
of (2) by γ0 (L) and ω (L), we get:

γ0 (L) c (L) ω (L) xt = γ0 (L) d (L) ω (L) yt∗ + γ0 (L) ω (L) ext ,
t = 0, 1, 2, ... (3)
2
The multivariate version where yt or xt or both are vectors, different lag lengths of the polynomials in
d(L) and c(L), and an MA component, can be easily handled, but at the cost of an additional complication
in the notation.

5
or

c0 (L) xt = g0 (Z)yt + γ
e e0 (L) ext , (4)
t = 0, 1, 2, ...

The specification in (4) is an example of an exact reverse unrestricted MIDAS model.

A few comments are in order. First, in (4) the HF variable x depends on its own HF lags,
on LF lags of the (observable) LF variable y, and on an error term that in general has an
M A structure.
Second, the model specification depends on the particular HF period we are in. Specif-
ically, for each i = 0, ..., k − 1, we can define a polynomial in the HF lag operator, γi (L),
such that the product gi (L) = γi (L)d(L) only contains powers of Lk−i , so that multiplying
both sides of (2) by γi (L) and ω (L), we get:

ci (L) xt = gi (Lk−i )yt + γ

e ei (L) ext , (5)
i i i
t = 0 + , 1 + , 2 + , ...
k k k
i = 0, ..., k − 1.

Therefore, RU-MIDAS models have a periodic structure.

Third, it can be shown that the polynomials γi (L) exist and can be analytically deter-
mined (see, e.g., Marcellino (1999) and the references therein).
Fourth, in practice the parameters of (2) are unknown so that γi (L) cannot be exactly
determined. Hence, empirically, we will use approximate reverse unrestricted MIDAS (RU-
MIDAS) models based on linear lag polynomials, such as

ai (L) xt = bi (Lk−i )yt + ξit ,

e (6)
i i i
t = 0 + , 1 + , 2 + , ...
k k k
i = 0, ..., k − 1.

ai (L) and bi (Lk−i ), ai and bi respectively, are large enough to make ξit
where the orders of e
white noise. Moreover, the RU-MIDAS model in (6) can be easily extended to allow for
additional HF or LF explanatory variables.
The estimation of the model is straightforward, since it is a linear model. Hence, for
each value of i the parameters of the model (6) can be estimated by OLS, and the lag
orders ai and bi can be selected with standard information criteria.

6
 Alternatively, given that the error terms ξit are in general correlated across i, the RU-
MIDAS equations for different values of i can be jointly estimated by means of a system
estimation method. This can be also achieved by grouping the equations in (6) into a single
equation with a proper set of dummy variables. As an example, consider the case where
k = 3 (e.g., monthly and quarterly variables), and ai = 3, bi = 1. The single equation
version of (6) is then

xt = α1 (1 − D2 − D3 ) yt− 1 + α2 D2 yt− 2 + α3 D3 yt−1 + (7)

3 3

β11 (1 − D2 − D3 ) xt− 1 + β12 D2 xt− 1 + β13 D3 xt− 1 +

3 3 3

β21 (1 − D2 − D3 ) xt− 2 + β22 D2 xt− 2 + β23 D3 xt− 2 +

3 3 3

β31 (1 − D2 − D3 ) xt−1 + β32 D2 xt−1 + β33 D3 xt−1 +

+vt ,
1 2
t = 0, , , 1, ...,
3 3
where D2 and D3 are dummy variables taking the value one in, respectively, each third and
first month of a quarter (so, respectively, for t = 32 , 1 + 23 , 2 + 32 , ... and t = 0, 1, 2, ...). The
model in (7) should be then estimated by GLS to allow for possible serial correlation and
heteroskedasticity.3
To conclude, a major advantage of the RU-MIDAS specifications in (6) is that they
are linear, so that standard techniques can be used for forecasting future values of the x
variable using information from their past values and the past values of the LF variable y.
The direct forecasting method (see, e.g., Marcellino et al. (2006)) is particularly useful in
this context, since it does not require to forecast future values of the y variable, although
it requires to change the model specification for each forecasting horizon.

3 Competing Models

An alternative approach to modeling the dependence of high-frequency on low-frequency

variables is based on a mixed frequency VAR (MF-VAR), using its two most popular for-
mulations. First, the typical MF-VAR, as in Mariano and Murasawa (2010), is cast in
3
We do not expect any major efficiency gains from the use of the system estimation approach with
respect to the univariate method, while model specification becomes more complex in the case of higher
frequency mismatch or several explanatory variables. Hence, in the Monte Carlo and empirical applications
we will focus on the univariate estimation approach.

7
state-space form, and the low-frequency series is treated as high-frequency series with miss-
ing observations. Second, the MF-VAR as described in Ghysels (2014), who decomposes
each high-frequency variable into a set of k low-frequency variables and models them in
low-frequency jointly with the low-frequency variable.

In the rest of this section, we provide details for both approaches and compare them to
the RU-MIDAS introduced in the previous section. We anticipate that a key advantage of
the RU-MIDAS is to provide an analytical relationship between high-frequency and low-
frequency variables, and a model that directly operates in high-frequency, emphasizing the
changing impact of the low-frequency variable on the high-frequency variable.

3.1 Mixed-Frequency VAR in state-space form

The mixed-frequency VAR cast in state-space form originates from the work of Mariano
and Murasawa (2010), and was subsequently used in a forecasting exercise by Kuzin et al.
(2011), see also Foroni et al. (2013a). Schorfheide and Song (2015) show how to estimate
this model in a Bayesian framework, thereby making the estimation of relatively large scale
models computationally feasible.

Focusing on the case studied by Mariano and Murasawa (2010), where yt is quarterly
GDP growth (a flow variable) and xt is a monthly variable, so that k = 3, yt is disaggregated
at the monthly frequency via the geometric mean:
1 2 ∗ 2 ∗ 1 ∗
yt = yt∗ + yt− 1 + y
∗
t− 32
+ yt−1 + yt− 4, (8)
3 3 3 3 3 3
where, as in the case of the RU-MIDAS model, yt∗ is the corresponding unobserved monthly
variable.

The latent variable yt∗ and the monthly indicator xt follow a bivariate VAR(p) process:

Φ (L) zt = ut , ut ∼ i.i.d.N (0, Σ) , (9)

where Φ (L) = I − Φ1 L − ... − Φp Lp and

" #
yt∗
zt = .
xt

8
Assuming that p ≤ 5, defining the state vector of unobserved variables st as
h i0
st = zt zt− 1 zt− 2 zt−1 zt− 4 ,
3 3 3

and given the relationship between yt and yt∗ in (8), the state-space representation of the
MF-VAR(p) is given by the following transition and measurement equations:4

st = Ast− 1 + Bvt (10)

3

" #
yt
= Cst , (11)
xt
where vt ∼ i.i.d.N (0, I2 ) and the system matrices A, B and C are given by:
" #
A1
A= , A1 = [Φ1 ...Φp 02∗2(5−p) ], A2 = [I8 08∗2 ],
A2
" #
Σ1/2
B= , C = [H0 ...H4 ],
08∗2
where the matrix C contains the coefficient matrices in the lag polynomial H(L) = 4i=0 Hi Li ,
P

which is defined according to the aggregation constraint defined in equation (8), and with
Lj xt = xt− j as in Section 2:
k

! ! ! ! !
1 2 2 1
0 0 1 0 2 0 3 0
H(L) = 3
+ 3
L+ L + 3
L + 3
L4 .
0 1 0 0 0 0 0 0 0 0

We refer to this model as the MF-VAR-KF.

In terms of parameter estimation for the MF-VAR-KF, the optimal procedure in this
context is Maximum Likelihood, which can be easily implemented. As in Mariano and
Murasawa (2010), we replace missing observations in the low-frequency variable with zeros
and rewrite the measurement equation accordingly, as if the missing values were random
draws from a Normal distribution N (0, 1), independent of model parameters:
" # " # " #
yt+ C1t D1t
= st + vt , (12)
xt C2 0
4
For brevity, we do not report here the representation for p > 5. Details for this specification can be
found in Mariano and Murasawa (2010). The state-space form can be easily modified to model stock rather
than flow LF variables, or to consider alternative aggregation schemes.

9
where vt is the random draw, and
(
C1 if yt is observable
C1t = . (13)
0 otherwise
(
0 if yt is observable
D1t = . (14)
I otherwise

For further details on the estimation, we refer to Mariano and Murasawa (2010). For the
VAR lag length choice, we use standard information criteria.

Moving to forecasting, from equation (10), the h-period-ahead forecast for the state
vector is given by:

ŝt+h|t = E[st+h |(yt , yt−1 , ..., y1 , xt , xt− 1 , ..., x1 )] = Ah ŝt|t , (15)

3

and the associated forecast error is

(st+h − ŝt+h|t ) = Ah (st − ŝt|t ) + Ah−1 vt+ 1 + Ah−2 vt+ 2 + ... + Avt+h− 1 + vt+h , (16)
3 3 3

so that the forecast error variance for the state vector is:

Pt+h|t = E[(st+h − ŝt+h|t )(st+h − ŝt+h|t )] = Ah Pt|t (Ah )0 + Ah−1 B(Ah−1 )0 + ... + B.

From equation (12), the h-period ahead forecast for the observable HF x variable, x̂t+h|t ,
can then be obtained from the following equation
" #
ŷt+h|t
ẑt+h|t = = Ct+h ŝt+h|t .
x̂t+h|t

Hence, the forecast error is the second element of the vector Ct+h (st+h − ŝt+h|t ), while the
forecast error variance is the element (2,2) of the matrix Ct+h Pt+h|t C0t+h .

Finally, to compare the MF-VAR-KF and the RU-MIDAS forecasts, we note that in
practice with the MF-VAR-KF we obtain the best estimates and forecasts for the HF
variable yt∗ and then we iterate forward the VAR in (9) and take expectations conditional
on the available information set to produce forecasts for xt . This is similar to using equation
(2), replacing y ∗ with interpolated values of y. Instead, in the (approximate) RU-MIDAS
model, we use the observable LF y as explanatory variable, which generates the periodic
structure of the model, and adopt a direct forecasting approach, which does not require the

10
specification of a model for y (and y ∗ ). Since the Kalman Filter produces the best linear
forecasts in the MSE sense, the MF-VAR-KF should dominate the RU-MIDAS if we know
the joint HF data generating process (DGP), that is, if the variables are generated by (9).
However, in practice the DGP is unknown, so that the comparison between MF-VAR-KF
and the RU-MIDAS is more complex and depends on the extent of the mis-specification
of the assumed joint HF DGP, which prevents the derivation of general theoretical results.
From the Monte Carlo simulations and empirical analyses reported in the following sections,
there is indeed no clear winner.

3.2 Mixed-Frequency VAR in stacked form

The mixed-frequency VAR in stacked form (MF-VAR-SF), developed by Ghysels (2014),

deals with the different data frequencies by stacking the HF variables in a LF vector of
dependent variables, depending on the timing of their releases. Continuing the case of the
previous subsection where there
 is only one quarterly variable y and one monthly variable
 0
x, the stacked vector is qt = xt− 2 , xt− 1 , xt , yt , where t = 1, 2, 3....
3 3

Let us assume that qt follows a VAR(p) process:

qt = A1 qt−1 + ... + Ap qt−p + et , (17)

where et is white noise with E (et e0t ) = V. The matrices Ai , i = 1, ..., p, and V are
subject to a set of restrictions when (17) is explicitly derived from a HF VAR such as
the one described by equation (9), see Ghysels (2014) for additional details. However,
typically these restrictions are not imposed to avoid dependence of the MF-VAR-SF from
an unknown HF DGP, as in the case of the RU-MIDAS. Given that there are no missing
observations in the system, this approach circumvents the estimation via the Kalman filter,
and standard least squares methods can be used to estimate the MF-VAR-SF. Lag length
selection can be based on information criteria.

Forecasting from this model seems simple, since it is a standard VAR. However, the
timing in (17) is in LF, t = 1, 2, 3..., while we may also want to forecast in HF, e.g., update
the forecast for x in each month of the quarter. This creates additional complications, and
the formula for the optimal forecast changes periodically. We illustrate this issue, assuming
that p = 1 and h = {1, 2, 3} to simplify the notation.

11
 The first possibility is that we want to forecast xt+1 without knowing xt+ 1 , xt+ 2 and
3 3
yt+1 , i.e., we are interested in xbt+1|t , which is a 3-step ahead forecast (or a 1-step ahead
forecast in LF). This is the standard case where one just uses the MF-VAR-SF model in
(17) and x bt+1|t coincides with the third element of the vector qbt+1|t , with

q
bt+1|t = A1 qt . (18)

Let us now assume that we have the same target, xt+1 , but we are in period t+ 13 , so that
we want to derive xbt+1|t+ 1 , which is a 2-step ahead forecast. In this case we need to use the
3
structural representation of the MF-VAR-SF, which describes the contemporaneous corre-
lations among the variables. Therefore, let us define P as the matrix of contemporaneous
correlations that is obtained via a Choleski decomposition
 
1 0 0 0
 p21 1 0 0 
P= , (19)
 
 p31 p32 1 0 
p41 p42 p43 1
0
with V = P−1 D (P−1 ) , where D is a diagonal matrix. Premultiplying (17) by P, we obtain
the structural MF-VAR-SF representation:

Pqt = PAqt−1 + ut , (20)

or, equivalently,
   
ux1
   
1 0 0 0 xt− 2 α11 α12 α13 α14 xt− 5 t
3 3

ux2
     
 p21 1 0 0  x 1
  t− 3  
= α21 α22 α23 α24  x 4
  t− 3  
+ t

, (21)
 
ux3
  
 p31 p32 1 0   xt  
 α31 α32 α33 α34   xt−1  
 t 
p41 p42 p43 1 yt α41 α42 α43 α44 yt−1 uyt

where [αij ] = [PA]ij is the matrix of structural coefficients, and ut = Pet is the vector of
structural errors.

If we focus on the dynamics of the x variable, we can see how it changes depending on
the month of the quarter. In the first month of the quarter (xt− 2 or xt−1+ 1 ), the dynamics
3 3
of xt is:
xt− 2 = α11 xt− 5 + α12 xt− 4 + α13 xt−1 + α14 yt−1 + ux1
t . (22)
3 3 3

12
In the second month of the quarter (xt− 1 or xt−1+ 2 ), the dynamics of xt becomes:
3 3

xt− 1 = α21 xt− 5 + α22 xt− 4 + α23 xt−1 − p21 xt− 2 + α24 yt−1 + ux2
t . (23)
3 3 3 3

In the third month (xt ), the dynamics of xt is:

xt = α31 xt− 5 + α32 xt− 4 + α33 xt−1 − p31 xt− 2 − p32 xt− 1 + α34 yt−1 + ux3
t . (24)
3 3 3 3

Therefore, our 2-step ahead forecast of interest x

bt+1|t+ 1 can be obtained as:
3

bt+1|t+ 1 = α31 xt− 2 + α32 xt− 1 + α33 xt + α34 yt − p31 xt+ 1 − p32 x
x bt+ 2 |t+ 1 , (25)
3 3 3 3 3 3

bt+ 2 |t+ 1 = α21 xt− 2 + α22 xt− 1 + α23 xt + α24 yt − p21 xt+ 1 .
x (26)
3 3 3 3 3

bt+1|t+ 2 (a 1-step ahead forecast), we should use the expression:

If instead we want x
3

bt+1|t+ 2 = α31 xt− 2 + α32 xt− 1 + α33 xt + α34 yt − p31 xt+ 1 − p32 xt+ 2 .
x (27)
3 3 3 3 3

All these expressions can be easily generalized to longer forecast horizons.

Ghysels (2014) compares in details the MF-VAR-SF and the MF-VAR-KF, which are
basically equivalent when the restrictions imposed by the HF VAR underlying the MF-
VAR-KF are imposed on the MF-VAR-SF. When instead the restrictions are not imposed,
the relative performance depends on the mis-specification of the HF DGP assumed by the
MF-VAR-KF and the goodness of fit for the MF-VAR-SF specification.

To conclude, at first sight, the MF-VAR-SF is rather different from the RU-MIDAS
approach. However, the detailed derivation of the optimal forecasts from the MF-VAR-SF
reveals that the two approaches are very similar, in the sense that also in the MF-VAR-SF
we need to change the specification of the HF variable model depending on the specific
month of the quarter, and the HF variable is regressed (in LF) on its own HF lags and on
LF lags of the LF variable. However, while the MF-VAR-SF is specified as a system, the
RU-MIDAS equations are specified one by one, which can give more flexibility. Moreover,
the forecasting approach is based on the direct method in the case of the RU-MIDAS and
on the iterated method for the MF-VAR-SF.

13
4 Monte Carlo Experiments

4.1 High-frequency DGP

In this section, we evaluate in a controlled experiment the forecasting ability of the

three different models outlined in the previous sections (RU-MIDAS, MF-VAR-KF, and
MF-VAR-SF models). The DGP we use is a bivariate model in the HF temporal unit
described as ! ! ! !
yt ρ δl yt−1 yt
= + , (28)
xt δh ρ xt−1 xt
where yt ∼ N (0, σ y ) and xt ∼ N (0, σ x ).

The design of the Monte Carlo experiment is as follows. First, time series with sample
size T = 600 are generated after discarding the first 100 observations to account for start-up
effects. The size of the evaluation sample is set to 150, and the estimation sample is recur-
sively expanded as we progress in the forecasting exercise, with the first estimation sample
extending over the first 450 observations.5 Second, for all models, we assume that the yt
variable is only observed once every three periods, so as to mimic the case of an empirically
relevant situation of a mix between quarterly and monthly variables. Third, for all mod-
els we calculate forecasts for the HF variable with forecast horizons h = {1, 2, 3, 6, 9, 12}
using an AR model as a benchmark. Fourth, the MF-VAR-SF model is estimated with-
out imposing any restrictions on the parameters of the autoregressive matrices to provide
a straightforward comparison with the RU-MIDAS model. The unrestricted version of
the MF-VAR-SF model is also easier to estimate, since standard least squares estimation
method can be used. In addition, results from Foroni et al. (2015) suggest that the forecast-
ing performance of the restricted and unrestricted versions are relatively similar. Finally,
for all models, the lag length is set so as to include one LF temporal unit of information. In
practice, this implies that the lag length for the MF-VAR-KF, RU-MIDAS and AR models
is set to 3, whereas the lag length in the MF-VAR-SF model is set to 1.

In addition, we use different values for the parameters when generating time series.
First, we assume a recursive DGP, setting δl = 0. The persistence parameter ρ is set to
5
Note that for computational reasons, the MF-VAR-KF model is estimated only once for each replication
(at the beginning of the forecasting exercise). Limited evaluation of a fully recursive forecasting exercise
produced qualitatively similar results.

14
0.8, a value that is relevant given the high persistence typically observed in macroeconomic
variables, and the parameter δh is set to 0.5. Second, we assume a non-recursive DGP,
using the following set of parameters (ρ, δh , δl ) = (0.5, 0.5, 0.4). These values are chosen so
as to make sure that the VAR satisfies (weak) stationarity conditions. Finally, the shocks yt
and xt are drawn independently from a normal distribution using the following parameter
values for the variance ({σ y , σ x } = {2, 1}).
We also calculate forecasts obtained with an Autoregressive Distributed Lag (ARDL)
model. This model is written as follows
p q
X X
∗
xt = α + βi xt− i + γj yt− j + ut , (29)
k k
i=1 j=1

where yt∗ is the high-frequency estimate of the low-frequency variable obtained by interpo-
lation. We use a linear interpolator, where the interpolated values for the LF variable are
obtained as follows
∗
yt+λ = (1 − λ)yt + λyt+1 , (30)
where λ = ki , i = {1, 2}, k = 3, yt+1 is the next non-missing value, and yt is the previous
non-missing value. This interpolation method requires the use of future values for the LF
variable in equation (30) (i.e., yt+1 ), which is problematic in a forecasting context. As
a result, in the forecasting exercise, the predicted values for the LF variable in equation
(30) are obtained from an AR model. Moreover, the lag length in equation (29) is set to
p = 3 and q = 3, so as to include one LF temporal unit of information to ensure a similar
information set across all models. Forecasts from ARDL models are calculated with the
direct method.
To evaluate the forecasts, we use the mean square prediction errors (MSPE) relative to
the MSPE calculated from the forecasts derived from an AR model. We report the median
of the estimates over 1000 replications. In doing so, we distinguish four different cases: (i)
MSPE calculated from the 150 observations of the evaluation sample, (ii) MSPE calculated
from the 50 observations where the forecast with horizon h = 1 corresponds to the first
month of the quarter (or first HF temporal unit of the LF temporal unit), (iii) MSPE
calculated from the 50 observations where the forecast with horizon h = 1 corresponds
to the second month of the quarter, and (iv) MSPE calculated from the 50 observations
where the forecast with horizon h = 1 corresponds to the third month of the quarter.
This is relevant given that the LF variable is observed periodically in our simulation set-up
(i.e., in the third month of each quarter). Hence, in the third month of the quarter, the

15
low-frequency variable has just been released, and it provides a valuable information when
forecasting the HF variable in the first month of the next quarter. As a result, it could
well be that for a given forecast horizon, the forecasting performance of the model varies
depending on the temporal distance to the release of the LF variable. Table 1 presents the
results when all observations of the evaluation sample are pooled when calculating MSPEs,
whereas Tables 2, 3 and 4 present the results when using only observations corresponding
to a specific month of the quarter when calculating the MSPEs.

The results for the high-frequency DGPs are presented in Panels A of these tables.
First, looking at Panel A of Table 1, the MF-VAR-SF model obtains the best forecasting
results for forecast horizons h = {1, 2}. Second, the RU-MIDAS model obtains the best
forecasting performance for h = {3, 6} in the case of the recursive DGP, and for h =
{3} in the case of the non-recursive DGP. Third, the MF-VAR-KF model typically ranks
closely to the MF-VAR-SF model. In fact, it obtains the best forecasting performance for
horizon h = {6, 9, 12} when the data are generated from a non-recursive DGP.6 Finally,
the ARDL model forecasting performance is somewhat inferior compared with the other
models, especially for fairly long forecast horizons (i.e., for h > 2).

Table 2 presents the results when the forecasts are evaluated in the first month of each
quarter, that is, right after the release of the LF variable. Unsurprisingly, the forecasting
performance of all models drastically improves. However, the ranking of the models is
little changed in that the MF-VAR-SF model continues to obtain the best forecasting
performance in the case of the recursive DGP, and it is closely followed by the MF-VAR-
KF and RU-MIDAS models. Table 3 shows the results when the forecasts are evaluated
in the second month of each quarter, that is, the forecasts are calculated one month after
the release of the LF variable. The forecasting performance of the models deteriorates
compared with Table 2. Interestingly, the MF-VAR-SF model continues to rank as the
best performing model for most forecast horizons with both the recursive and non-recursive
DGPs. Table 4 shows that the forecasting performance of all models further deteriorates
when forecasts are calculated two months after the release of the LF variable. In fact, in
nearly all cases, the forecasting performance of the AR model is better or comparable to
that of the mixed-frequency models.
6
In the appendix, we also consider a different aggregation rule for the MF-VAR-KF model than the
one described by equation (8). We find that, conditional on our simulation setup, there is no substantial
forecasting gain to expect from the use of a different aggregation rule.

16
4.2 Mixed-frequency DGP

One caveat on the simulations presented so far is that data are generated from a high-
frequency VAR, which may not be realistic given that macroeconomic time series are typi-
cally sampled at different frequencies. As a result, we evaluate here the forecasting ability
of the different models using mixed-frequency DGPs, generating data from a MF-VAR-SF
model.

In line with simulations presented in Götz and Hecq (2014), data generated from a
MF-VAR-SF are obtained from the following dynamic structural equations:
     
ux1
 
1 0 0 0 xt− 2 0.6 −0.2 0.4 0.3 xt− 5 t
3 3

 ux2
 −0.4     
1 0 0   x t− 1  0 0.6 −0.2 0.3  xt− 4   t 

=  +  x3  .
  3     3 
  
 0.2 −0.4 1 0    xt   0 0 0.6 0.3   xt−1   ut 
0 0 δ 1 yt 0 0 0 0.5 yt−1 uyt

We choose two different values for the parameter δ that determines the contemporaneous
correlation between the LF and HF variable (δ = 0 and δ = 0.2).7 Note that while data
are not directly generated from a RU-MIDAS model, setting δ = 0 roughly corresponds to
the case of a RU-MIDAS DGP. Apart from generating time series from different DGPs, the
design of the Monte Carlo exercise is identical to the one presented in the previous section.

Panels B of Tables 1 to 4 present the results. As expected, the MF-VAR-SF model

(i.e., the model used to generate the data) obtains in nearly all cases the best forecasting
performance. It is also interesting to note that the RU-MIDAS model exhibits a very similar
forecasting performance to the MF-VAR-SF model for forecast horizons h = {3, 6, 9, 12},
a pattern that could be also observed when data were generated from a high-frequency
DGP. While the ranking of the models is somewhat similar across the different types of
DGP (high-frequency or mixed-frequency DGPs), when data are generated from the mixed-
frequency DGP, we no longer observe a clear trend towards a better forecasting performance
depending on the distance to the release of the LF variable. This suggests that the relevance
of calculating the forecasting performance in a periodic way varies depending on the data
generating process. Hence, it remains ultimately an empirical issue that we investigate in
the next section.
7
Note that when the LF and HF variables are heavily correlated, a multicollinearity problem may
arise, which heavily deteriorates the forecasting performance of the RU-MIDAS model at selected forecast
horizons. In such cases, appropriate transformation of the HF variable permits to alleviate this issue.

17
 Overall, the Monte-Carlo results suggest that, conditional on our simulation set-up,
the forecasting performance may vary considerably depending on the temporal distance to
the release of the LF variable. In terms of ranking of the models, our simulation exercise
suggests that there is no single outperforming model. However, the MF-VAR-SF model
tends to outperform the competing models, and the RU-MIDAS model performs largely in
line with the MF-VAR-SF model at longer forecasting horizons.

5 Empirical Applications

It is well documented in the literature that surveys are helpful indicators to predict key
macroeconomic variables (see, for example, Chun (2011), Chernov and Mueller (2012) and
Stark (2010)). For the U.S. economy, it is typically difficult to improve upon the forecasting
performance of well established surveys such as the Survey of Professional Forecasters
(see, e.g., Faust and Wright (2013)). In a similar vein, survey data can also be used in
standard macroeconomic forecasting models. For example, Wright (2013) shows that using
survey data to calibrate prior distributions in Bayesian VAR models yields a substantial
improvement in real-time macroeconomic predictions compared with models that exclude
such information.

To illustrate the forecasting ability of the different mixed-frequency data models pre-
sented in the previous sections, we exploit the natural frequency mismatch that exists
between some widely used survey data (typically available on a quarterly frequency) and
the target variable (typically sampled at a monthly frequency). In doing so, we use the
Survey of Professional Forecasters (SPF) from the Federal Reserve Bank of Philadelphia,
which is widely recognized as a very accurate survey (see, e.g., Stark (2010), who performs
an extensive evaluation of the SPF forecasting performance).

We run two different empirical exercises. First, we illustrate the forecasting performance
of the different models using the quarterly SPF for three key monthly variables represen-
tative of the U.S. economy: CPI inflation, industrial production growth, and personal
consumption growth. Our goal is to check whether using the information from quarterly
SPF data is helpful to improve the forecasts of the corresponding monthly series. Second,
we forecast the monthly nominal U.S. dollar trade-weighted exchange rate, which is a chal-
lenging task given that the random walk model is typically found to be the best performing

18
model when forecasting exchange rates, see e.g. Meese and Rogoff (1983a) for early evi-
dence and Engel and West (2005) for a theoretical explanation of this finding. Again, the
objective is to evaluate whether the information conveyed by the quarterly SPF helps to
forecast the monthly U.S. trade-weighted exchange rate.

5.1 Design of the forecasting experiments

We apply the following transformation to the monthly variables. Industrial production

growth, consumption growth and inflation rate are the annualized three-month percentage
change (i.e., 400 times the three-month log change in the level of the industrial production
index, real personal consumption expenditures index and the CPI index). Moreover, the
exchange rate measure we use is the nominal trade-weighted U.S. exchange rate (i.e., a
weighted average of the foreign exchange value of the U.S. dollar against major currencies),
and it is taken in logs. All the data are downloaded from the ALFRED real-time database
of the Saint Louis Federal Reserve Bank. The full sample size for all time series extends
from July 1981 to November 2013 and we use real-time data, whenever available.

For the quarterly variables, we consider the median forecasts for the CPI inflation
rate, industrial production and personal consumption. The SPF survey for inflation and
consumption growth starts in the third quarter of 1981, which explains the start date of
our sample. For ease of comparison across models and variables, the estimation sample
for industrial production also starts in 1981Q3 (even though the SPF for this variable is
available from 1968Q4).

Figure 1 reports the monthly series together with the expectations derived from the SPF
referring to the current quarter and one-quarter ahead. This shows that the SPF tracks
relatively well the actual realizations of the monthly macroeconomic variables, albeit the
SPF is less volatile than the actual underlying series.

The first estimation sample extends from July 1981 to December 1999, and it is re-
cursively expanded until September 2011. For each month of the evaluation sample, we
calculate forecasts from 1 to 24 months ahead, and the forecasts are evaluated against the
first release, calculating the Mean Square Prediction Error (MSPE). We consider an AR
model as a benchmark with lag length selected according to the BIC criterion (and a max-
imum lag equal to 12). The forecasting performance of a given model is then presented as
a ratio with respect to the MSPE obtained from the AR benchmark. When forecasting the

19
monthly trade-weighted exchange rate, we use the random walk as benchmark model, since
it is the traditional benchmark model in this literature (see Rossi (2013)). Note also that
in line with the Monte-Carlo experiments, in the forecasting exercise, both the MF-VAR-
KF and MF-VAR-SF models are estimated once a quarter, but the (iterated) forecasts are
updated on a monthly basis so as to take into account the latest information. However,
both ARDL model and RU-MIDAS models are estimated on a monthly basis, since the
forecasts are calculated with the direct method so that the parameter estimates change
with the forecast horizon.

In total, for each forecast horizon and model, we obtain 144 forecasts. In line with the
Monte Carlo experiments, we report the MSPE calculated when pooling all forecasts from
the evaluation sample, but also report the MSPE calculated in the first, second and third
month of each quarter. This is relevant since we are looking at periodic models, in which
the interaction between the monthly and quarterly variables changes for every month of
the quarter.

Moreover, the SPF provides forecasts for different quarterly horizons. For each of the
variables, we can consider five series of forecasts: the series of forecasts related to the current
quarter, and the series of forecasts from 1 to 4-quarter ahead. However, when reporting
the relative MSPE, we use the SPF related to the specific forecast horizon. In detail, for
forecast horizon h = {1, 2, 3}, we use the SPF related to 1-quarter ahead forecast, for
forecast horizon h = {6}, we use the SPF related to 2-quarter ahead forecast, for forecast
horizon h = {9}, we use the SPF related to 3-quarter ahead forecast, and for forecast
horizons h = {12, 15, 18, 21, 24}, we use the SPF related to 4-quarter ahead forecasts.

5.2 Predicting monthly macroeconomic variables

In this subsection, we assess the relevance of the SPF to forecast the corresponding
monthly variables, based on competing low to high frequency models. In Tables 5 to 7 we
present the MSPE for each of our models relative to the AR benchmark. The tables report
the ratio of the MSPE computed on the entire evaluation sample, but also separate results
for the first, second and third month of each quarter. The numbers in bold highlight the
cases in which the mixed-frequency models outperform the benchmark (i.e., when the ratio
of the MSPE of the mixed-frequency model relative to the MSPE of the AR benchmark
is smaller than 1). In addition, we implement the Diebold and Mariano test to formally

20
test for significant differences in the forecasting performance between a given model and
the AR benchmark model.

Table 5 shows that SPF expectations on inflation contain useful information for predict-
ing monthly inflation. Using information included in the SPF allows us to improve upon
the AR benchmark at almost every horizon (with the exception of the very short ones)
for all the models we analyse. In a substantial number of cases, the improvement is also
statistically significant according to the Diebold and Mariano test. In detail, for short fore-
cast horizons (h={1,2,3}), the ARDL model based on interpolated survey data performs
particularly well. At longer horizons (i.e., for h > 3), the more sophisticated models have
an advantage, with the MF-VAR-SF often ranking as best and the RU-MIDAS as second
best. Results are generally somewhat better in the third month of the quarter. This hints
at the fact that useful information is contained in the new release of the SPF, which are
available in the middle of the second month of each quarter, and consequently enter the
forecast in the third month in a timely way. The gains are overall rather stable across the
three months of the quarter, in the range of 10 to 20 per cent, and are often statistically
significant.

We find similar results when forecasting monthly industrial production growth (see
Table 6). The gains from the use of the corresponding quarterly surveys are evident, for
any forecast horizon. In particular, good results are obtained in the third month of the
quarter, when the new releases of the SPF become available. In line with the results for
forecasting inflation, at very short horizons (h = {1, 2}), the ARDL provides the most
accurate forecasts. The other models are typically better for h > 2, with the differences
across models generally small and gains with respect to the AR benchmark between 5%
and 10% and often significant.

The information contained in the SPF appears to be less useful for predicting per-
sonal consumption growth (see Table 7). Looking at Figure 1, it is clear that personal
consumption growth is more volatile than the other two series we consider, and than the
corresponding SPF. Therefore, in this case, survey data do not convey useful information
especially at short horizons. However, the SPF seems to have some predictive content for
the longer horizons, with some gains for h ≥ 6. The best models in this case are the
MF-VARs, which obtain gains close to 5%.

21
5.3 Predicting the monthly U.S. trade-weighted exchange rate

As a second forecasting exercise, we now focus on predicting exchange rates. In particu-

lar, we investigate the predictive power of the expectations on future inflation and interest
rates for the monthly nominal U.S. trade-weighted exchange rate.8 The rationale for this
choice of predictors is that using expected inflation relates to the purchasing power parity
(PPP) theory, whereas interest rates have predictive power for exchange rate according to
the uncovered interest rate parity (UIP) relation. While both UIP and PPP refer to con-
temporaneous values for inflation and interest rate, using expectations for both variables is
likely to be more relevant in a forecasting context. Note also that, according to both UIP
and PPP, differentials between U.S. and foreign variables should be the relevant metrics to
look at for exchange rate fluctuations. However, expectations data for other countries are
limited in that they do not cover a long enough sample. As a result, we only use U.S. data
as regressors.9

Following the literature (see Rossi (2013) for an extensive review), in estimating our
models we focus on Et (st+h − st ), where st indicates the log of the exchange rate, with t
expressed in months and h the monthly forecast horizon. All models are evaluated relative
to a benchmark random walk without drift. The predictive ability of our models is evaluated
based on the mean squared prediction error.

The results are then presented in the same way as in the previous subsection. Tables
8 and 9 present the results when using the expectations on CPI inflation and on T-Bill
yields, respectively. Note also we use the SPF referring to the current quarter for all
the forecasting horizons, since, contrary to the previous subsection, there is no economic
rationale for doing otherwise. Interestingly, mixed-frequency data models outperform the
benchmark random walk model.

In particular, Table 8 shows that SPF expectations on inflation contain useful informa-
tion for predicting the monthly nominal trade-weighted U.S. exchange rate. Improvements
8
In detail, this is a weighted average of the foreign exchange value of the U.S. dollar against a subset
of the broad index currencies that circulate widely outside the country of issue. Major currencies index
includes the Euro Area, Canada, Japan, United Kingdom, Switzerland, Australia, and Sweden.
9
For the same reason, we focus on the trade-weighted exchange rate and not on bilateral exchange
rates, in which expectations for both countries would be relevant. Note that results based on out-of-sample
forecasting of selected currency pairs (GBP/USD, JPY/USD, and CAD/USD) only produced marginal
improvements in forecast accuracy for selected horizons. For the use of survey data for explaining bilateral
exchange rate fluctuations, see also Fratzscher et al. (2015).

22
are evident with every model we consider, and at every horizon. The gains are even in the
range of 10 to more than 15 percent, which are substantial in the context of exchange rate
forecasting, in which it is difficult to beat the no-change (or random walk) benchmark. Ad-
mittedly, results are not always statistically significant, except for the results obtained with
the MF-VAR estimated via the Kalman filter. This method provides better and significant
results in almost every case.10

Table 9 shows that expectations on the T-Bill seem to be a useful predictor only when
used in the MF-VAR setup, and especially with the MF-VAR estimated with the Kalman
filter. RU-MIDAS and interpolation methods do not provide good results.

All in all, we obtain encouraging results by using surveys of professional forecasters to

forecast the U.S. trade-weighted exchange rate in a mixed-frequency setup. In many cases,
mixed-frequency models obtain better forecasting results than the random walk benchmark,
and the forecasting gains are often statistically significant.

6 Conclusions

In this paper, we analyze how to incorporate low-frequency information in models that

forecast high-frequency variables. While the literature has mostly concentrated on the
use of high-frequency variables for predicting a lower frequency variable (e.g., forecasting
quarterly GDP growth with monthly indicators), little work has been done regarding the
use of low-frequency variables to predict higher frequency variables.

First, we introduce a new model, denoted as RU-MIDAS, in which the high-frequency

variable has a dynamic relation with the low-frequency variable. We then compare this
model with two different versions of the mixed-frequency VAR model (the Kalman filter
and stacked-form versions), so far only used to exploit high frequency information to predict
low frequency variables.

Second, we evaluate the forecasting ability of these competing mixed-frequency models,

by means of a number of Monte Carlo experiments. Our simulation results suggest that the
predictive performance of the low-frequency variable for a related high-frequency variable
increases soon after the release of the low-frequency variable, suggesting that the timing
10
This is consistent with the results from Dal Bianco et al. (2012), who use the MF-VAR-KF model to
forecast the euro-dollar exchange rate at the weekly frequency.

23
of the release of the low-frequency variable matters when predicting the high-frequency
variable. Moreover, across the different simulations we perform, no clear-cut ranking of the
competing mixed-frequency models emerges.

Finally, we illustrate the empirical relevance of the different mixed-frequency models

using the quarterly Survey of Professional Forecasters for forecasting three key US monthly
macroeconomic indicators (industrial production, consumption and inflation) as well as
the monthly nominal U.S. trade-weighted exchange rate. In line with the Monte Carlo
experiments, we do not find a clear-cut ranking in the forecasting performance of the
models. However, our results clearly indicate that there is additional insight to gain from
the quarterly SPF right after its publication when one is interested in forecasting the
corresponding monthly indicator. Hence, low frequency information can indeed be useful
when predicting high frequency variables.

24
References
Aruoba, S. B., Diebold, F. X., and Scotti, C. (2009). Real-Time Measurement of Business
Conditions. Journal of Business & Economic Statistics, 27(4):417–427.

Blasques, F., Koopman, S. J., , and Mallee, M. (2014). Low Frequency and Weighted
Likelihood Solutions for Mixed Frequency Dynamic Factor Models. Tinbergen Institute
Discussion Papers, 14-105/III.

Chernov, M. and Mueller, P. (2012). The term structure of inflation expectations. Journal
of Financial Economics, 106(2):367–394.

Chun, A. L. (2011). Expectations, Bond Yields, and Monetary Policy. Review of Financial
Studies, 24(1):208–247.

Clements, M. P. and Galvao, A. B. (2008). Macroeconomic Forecasting With Mixed-

Frequency Data. Journal of Business & Economic Statistics, 26:546–554.

Clements, M. P. and Galvao, A. B. (2009). Forecasting US output growth using lead-

ing indicators: an appraisal using MIDAS models. Journal of Applied Econometrics,
24(7):1187–1206.

Dal Bianco, M., Camacho, M., and Perez Quiros, G. (2012). Short-run forecasting of the
euro-dollar exchange rate with economic fundamentals. Journal of International Money
and Finance, 31(2):377–396.

Engel, C. and West, K. D. (2005). Exchange Rates and Fundamentals. Journal of Political
Economy, 113(3):485–517.

Eraker, B., Chiu, C. W. J., Foerster, A. T., Kim, T. B., and Seoane, H. D. (2015). Bayesian
Mixed Frequency VARs. Journal of Financial Econometrics, 13(3):698–721.

Faust, J. and Wright, J. H. (2013). Forecasting Inflation. Handbook of Economic Forecast-

ing, 2(A):1063–1119.

Foroni, C., Ghysels, E., and Marcellino, M. (2013a). Mixed-frequency vector autoregressive
models. Advances in Econometrics, 32:247–272.

Foroni, C., Guérin, P., and Marcellino, M. (2015). Markov-switching mixed-frequency VAR
models. International Journal of Forecasting, 31(3):692–711.

25
Foroni, C., Marcellino, M., and Schumacher, C. (2013b). U-MIDAS: MIDAS regressions
with unrestricted lag polynomials. Journal of the Royal Statistical Society - Series A,
forthcoming(8828).

Fratzscher, M., Rime, D., Sarno, L., and Zinna, G. (2015). The scapegoat theory of
exchange rates: the first tests. Journal of Monetary Economics, 70(C):1–21.

Ghysels, E. (2014). Macroeconomics and the reality of mixed-frequency data. Journal of

Econometrics, forthcoming.

Ghysels, E., Santa-Clara, P., and Valkanov, R. (2006). Predicting volatility: getting the
most out of return data sampled at different frequencies. Journal of Econometrics, 131(1-
2):59–95.

Ghysels, E. and Wright, J. H. (2009). Forecasting Professional Forecasters. Journal of

Business & Economic Statistics, 27(4):504–516.

Giannone, D., Reichlin, L., and Small, D. (2008). Nowcasting: The real-time informational
content of macroeconomic data. Journal of Monetary Economics, 55(4):665–676.

Götz, T. B. and Hecq, A. (2014). Nowcasting causality in mixed frequency vector autore-
gressive models. Economics Letters, 122(1):74–78.

Kuzin, V., Marcellino, M., and Schumacher, C. (2011). MIDAS vs. mixed-frequency VAR:
Nowcasting GDP in the euro area. International Journal of Forecasting, 27(2):529–542.

Marcellino, M. (1999). Some Consequences of Temporal Aggregation in Empirical Analysis.

Journal of Business & Economic Statistics, 17(1):129–36.

Marcellino, M., Stock, J. H., and Watson, M. W. (2006). A comparison of direct and
iterated multistep AR methods for forecasting macroeconomic time series. Journal of
Econometrics, 135(1-2):499–526.

Mariano, R. S. and Murasawa, Y. (2003). A new coincident index of business cycles based
on monthly and quarterly series. Journal of Applied Econometrics, 18(4):427–443.

Mariano, R. S. and Murasawa, Y. (2010). A Coincident Index, Common Factors, and

Monthly Real GDP. Oxford Bulletin of Economics and Statistics, 72(1):27–46.

26
Meese, R. and Rogoff, K. (1983a). The Out-of-Sample Failure of Empirical Exchange
Rate Models: Sampling Error or Misspecification? In Exchange Rates and International
Macroeconomics, NBER Chapters, pages 67–112. National Bureau of Economic Research,
Inc.

Meese, R. A. and Rogoff, K. (1983b). Empirical exchange rate models of the seventies :
Do they fit out of sample? Journal of International Economics, 14(1-2):3–24.

Rime, D., Sarno, L., and Sojli, E. (2010). Exchange rate forecasting, order flow and
macroeconomic information. Journal of International Economics, 80(1):72–88.

Rossi, B. (2013). Exchange rate predictability. Journal of Economic Literature, 51(4):2–56.

Schorfheide, F. and Song, D. (2015). Real-Time Forecasting with a Mixed-Frequency VAR.

Journal of Business & Economic Statistics, 33(3):366–380.

Stark, T. (2010). Realistic evaluation of real-time forecasts in the Survey of Professional

Forecasters. Research Rap Special Report, Federal Reserve Bank of Philadelphia, (May).

Wright, J. H. (2013). Evaluating Real-Time Var Forecasts With An Informative Democratic

Prior. Journal of Applied Econometrics, 28(5):762–776.

27
7 Appendix

Comparison of different aggregation rules for the MF-VAR-KF model

In this section, we perform additional simulations to check the relevance of the aggre-
gation constraint from Mariano and Murasawa (2003) when estimating the MF-VAR-KF
model (i.e., equation (8) in Section 3.1). This aggregation constraint is derived from a ge-
ometric mean, assuming that the (flow) data on a given quarter is three times its monthly
values pertaining to that quarter. However, this type of constraint is not appropriate in
the context of a stock variable where the aggregation rule is straightforward since stock
variables are just a particular quantity at a specific time. Also, the MF-VAR-KF model
is not directly comparable with the MF-VAR-SF and RU-MIDAS models to the extent
that there is no such disaggregation of the LF variable. Instead, in both RU-MIDAS and
MF-VAR-SF models, the LF variable is assumed to be observed every k periods. Hence,
the comparison of the MF-VAR-KF model with the RU-MIDAS and MF-VAR-SF models
may be distorted by the different aggregation rules adopted in these different models. As
a result, we consider a stock variable type of aggregation so that equation (8) becomes:

! ! ! ! !
1 0 0 0 0 0 0 0 0 0
H(L) = + L+ L2 + L3 + L4 (31)
0 1 0 0 0 0 0 0 0 0

if the LF variable is observed (otherwise, the upper row of H(L) is set to 0).

Table 10 shows the results. First, in the case of data generated from a high-frequency
VAR, the forecasting performance of the MF-VAR-KF models is very close, regardless
of the aggregation rule adopted. However, in the case of data generated from a mixed-
frequency VAR DGP, the results suggest that the MF-VAR-KF model with an aggregation
rule for stock variable outperforms the MF-VAR-KF model with an aggregation rule à la
Mariano and Murasawa (2003) for forecast horizon h={1}. However, at longer forecast
horizons, both models exhibit a similar forecasting performance. Overall, conditional on
this simulation exercise, this suggests that there is no substantial forecasting gain to expect
from using a different aggregation rule for the MF-VAR-KF model.

28
 Figure 1: Macroeconomic variables and SPF expectations

Inflation
10

-5
Actual data
-10 1-quarter ahead SPF
Current quarter SPF

1985 1990 1995 2000 2005 2010

Industrial Production

10

-10
Actual data
1-quarter ahead SPF
-20 Current quarter SPF

1985 1990 1995 2000 2005 2010

Real PCE

10

0 Actual data
1-quarter ahead SPF
Current quarter SPF
-5
1985 1990 1995 2000 2005 2010

Note: All macroeconomic variables are taken as 400 times the three-month change in the logarithm of the
underlying index.

29
 Table 1: Simulation Results - All months

Panel A: High-frequency DGP

Forec. hor. 1 2 3 6 9 12 1 2 3 6 9 12

DGP: (ρ, δh , δl , σ x , σ y ) = (0.8, 0.5, 0, 2, 1) (ρ, δh , δl , σ x , σ y ) = (0.5, 0.5, 0.4, 2, 1)

RU-MIDAS 0.859 0.914 0.804 0.915 0.983 1.005 0.946 0.979 0.934 0.987 1.006 1.007
MF-VAR-SF 0.771 0.771 0.809 0.915 0.968 0.990 0.827 0.898 0.935 0.978 0.986 0.989
MF-VAR-KF 0.794 0.792 0.832 0.932 0.989 1.006 0.881 0.917 0.941 0.971 0.974 0.971
ARDL 0.794 0.823 0.876 0.965 1.005 1.011 0.860 0.943 0.994 1.013 1.012 1.009

Panel B: Mixed-frequency DGP

Forec. hor. 1 2 3 6 9 12 1 2 3 6 9 12

DGP: MF-VAR-SF DGP (δ = 0) MF-VAR-SF DGP (δ = 0.2)

RU-MIDAS 0.794 0.792 0.799 0.830 0.874 0.929 0.801 0.816 0.805 0.847 0.895 0.946
MF-VAR-SF 0.685 0.672 0.797 0.819 0.858 0.907 0.684 0.675 0.801 0.834 0.878 0.922
MF-VAR-KF 0.927 0.877 0.826 0.874 0.936 0.990 0.931 0.897 0.831 0.896 0.950 0.998
ARDL 0.724 0.678 0.819 0.832 0.871 0.915 0.722 0.683 0.835 0.856 0.898 0.935

Note: This table reports the median of the relative Mean Square Prediction Error for the RU-MIDAS,
MF-VAR-KF, MF-VAR-SF and ARDL models averaged over 1000 replications. The benchmark model is
an AR model. Boldface indicates the model with the lowest relative MSPE for a given horizon and DGP.
Additional details on the DGPs are provided in the text.

30
 Table 2: Simulation Results - First month of the quarter

Panel A: High-frequency DGP

Forec. hor. 1 2 3 6 9 12 1 2 3 6 9 12

DGP (ρ, δh , δl , σ x , σ y ) = (0.8, 0.5, 0, 2, 1) (ρ, δh , δl , σ x , σ y ) = (0.5, 0.5, 0.4, 2, 1)

RU-MIDAS 0.552 0.548 0.528 0.792 0.923 0.992 0.661 0.714 0.788 0.910 0.961 0.981
MF-VAR-SF 0.432 0.432 0.528 0.787 0.914 0.969 0.483 0.683 0.788 0.898 0.947 0.967
MF-VAR-KF 0.440 0.445 0.543 0.797 0.926 0.974 0.503 0.679 0.794 0.904 0.934 0.957
ARDL 0.436 0.453 0.584 0.838 0.953 0.998 0.489 0.710 0.843 0.932 0.965 0.985

Panel B: Mixed-frequency DGP

Forec. hor. 1 2 3 6 9 12 1 2 3 6 9 12

DGP: MF-VAR-SF DGP (δ = 0) MF-VAR-SF DGP (δ = 0.2)

RU-MIDAS 0.912 0.811 0.564 0.628 0.758 0.863 0.943 0.847 0.568 0.656 0.803 0.898
MF-VAR-SF 0.716 0.559 0.563 0.625 0.751 0.842 0.718 0.556 0.569 0.652 0.790 0.875
MF-VAR-KF 0.741 0.568 0.590 0.651 0.780 0.889 0.757 0.563 0.596 0.685 0.827 0.923
ARDL 0.768 0.551 0.575 0.639 0.761 0.853 0.775 0.549 0.582 0.667 0.802 0.888

Note: This table reports the median of the relative Mean Square Prediction Error for the RU-MIDAS,
MF-VAR-KF, MF-VAR-SF and ARDL models averaged over 1000 replications. This table reports the
results when the MSPEs are calculated only from the months where the forecast with horizon h = 1 refers
to the first month of the quarter, that is one month after the LF variable has been released. The benchmark
model is an AR model. Boldface indicates the model with the lowest relative MSPE for a given horizon
and DGP. Additional details on the DGPs are provided in the text.

31
 Table 3: Simulation Results - Second month of the quarter

Panel A: High-frequency DGP

Forec. hor. 1 2 3 6 9 12 1 2 3 6 9 12

DGP (ρ, δh , δl , σ x , σ y ) = (0.8, 0.5, 0, 2, 1) (ρ, δh , δl , σ x , σ y ) = (0.5, 0.5, 0.4, 2, 1)

RU-MIDAS 0.893 0.936 0.875 0.953 0.994 1.007 1.068 1.085 0.986 1.015 1.018 1.011
MF-VAR-SF 0.864 0.859 0.882 0.949 0.983 0.988 0.962 0.988 0.986 0.996 1.002 1.001
MF-VAR-KF 0.940 0.918 0.914 0.972 1.007 1.006 1.046 1.015 1.000 1.001 1.004 0.979
ARDL 0.864 0.863 0.888 0.951 0.986 0.993 0.972 0.999 1.004 1.017 1.017 1.006

Panel B: Mixed-frequency DGP

Forec. hor. 1 2 3 6 9 12 1 2 3 6 9 12

DGP: MF-VAR-SF DGP (δ = 0) MF-VAR-SF DGP (δ = 0.2)

RU-MIDAS 0.750 0.620 0.881 0.883 0.895 0.923 0.747 0.623 0.886 0.908 0.935 0.949
MF-VAR-SF 0.641 0.600 0.877 0.871 0.886 0.907 0.639 0.600 0.885 0.896 0.911 0.928
MF-VAR-KF 0.972 0.817 0.894 0.931 0.967 0.997 0.985 0.830 0.908 0.957 0.994 1.028
ARDL 0.683 0.605 0.886 0.863 0.875 0.912 0.681 0.603 0.905 0.895 0.909 0.935

Note: This table reports the median of the relative Mean Square Prediction Error for the RU-MIDAS,
MF-VAR-KF, MF-VAR-SF and ARDL models averaged over 1000 replications. This table reports the
results when the MSPEs are calculated only from the months where the forecast with horizon h = 1 refers
to the second month of the quarter, that is two months after the LF variable has been released. The
benchmark model is an AR model. Boldface indicates the model with the lowest relative MSPE for a given
horizon and DGP. Additional details on the DGPs are provided in the text.

32
 Table 4: Simulation Results - Third month of the quarter

Panel A: High-frequency DGP

Forec. hor. 1 2 3 6 9 12 1 2 3 6 9 12

DGP (ρ, δh , δl , σ x , σ y ) = (0.8, 0.5, 0, 2, 1) (ρ, δh , δl , σ x , σ y ) = (0.5, 0.5, 0.4, 2, 1)

RU-MIDAS 1.135 1.248 0.997 1.011 1.020 1.029 1.095 1.155 1.012 1.020 1.019 1.026
MF-VAR-SF 1.011 1.008 1.006 1.011 1.005 1.012 1.030 1.020 1.011 1.010 1.008 1.010
MF-VAR-KF 1.018 1.020 1.015 1.031 1.016 1.008 1.066 1.046 1.007 1.005 0.989 0.989
ARDL 1.096 1.152 1.154 1.100 1.052 1.039 1.120 1.140 1.131 1.078 1.043 1.038

Panel B: Mixed-frequency DGP

Forec. hor. 1 2 3 6 9 12 1 2 3 6 9 12

DGP: MF-VAR-SF DGP (δ = 0) MF-VAR-SF DGP (δ = 0.2)

RU-MIDAS 0.732 0.944 0.939 0.954 0.971 0.976 0.995 1.019 1.007 1.015 1.022 1.019
MF-VAR-SF 0.691 0.859 0.942 0.950 0.954 0.951 0.944 0.942 0.997 0.992 0.997 0.996
MF-VAR-KF 1.050 1.259 0.980 1.034 1.062 1.064 1.076 1.281 0.975 1.032 1.038 1.039
ARDL 0.732 0.878 0.972 0.975 0.975 0.974 0.960 0.905 1.179 1.104 1.050 1.039

Note: This table reports the median of the relative Mean Square Prediction Error for the RU-MIDAS,
MF-VAR-KF, MF-VAR-SF and ARDL models averaged over 1000 replications. This table reports the
results when the MSPEs are calculated only from the months where the forecast with horizon h = 1 refers
to the third month of the quarter, that is three months after the LF variable has been released. The
benchmark model is an AR model. Boldface indicates the model with the lowest relative MSPE for a given
horizon and DGP. Additional details on the DGPs are provided in the text.

33
 Table 5: Forecasting monthly inflation with the quarterly SPF

All months pooled

Forec. hor. 1 2 3 6 9 12 15 18 21 24

RU-MIDAS 1.175 1.001 0.915 0.865* 0.919 0.946 0.865* 0.921 0.953 0.951
MF-VAR-SF 1.218 1.079 0.908 0.819** 0.811** 0.871* 0.857** 0.890** 0.904** 0.904**
MF-VAR-KF 1.783 1.299 1.015 0.890 0.868* 0.929 0.918* 0.958 0.973 0.981
ARDL 0.906* 0.874* 0.845* 0.882** 0.891* 0.966 0.958 0.938 0.949 0.961

First month

Forec. hor. 1 2 3 6 9 12 15 18 21 24

RU-MIDAS 1.374 0.939 0.944* 0.950 0.955 1.063 0.898** 0.940** 0.908** 0.979
MF-VAR-SF 1.323 1.046 0.967 0.872* 0.857* 0.924 0.886* 0.917* 0.908** 0.924**
MF-VAR-KF 1.394 1.308 1.170 0.956 0.902 0.981 0.940 0.974 0.970 0.995
ARDL 0.901* 0.915 0.923 0.916 0.919 0.985 0.955 0.958 0.931 0.896*

Second month

Forec. hor. 1 2 3 6 9 12 15 18 21 24

RU-MIDAS 1.101 1.004 0.897* 0.863* 0.852** 0.851** 0.874** 0.912** 1.041 0.965**
MF-VAR-SF 1.243 1.166 0.900 0.826 0.812* 0.857 0.877* 0.886* 0.921* 0.939*
MF-VAR-KF 2.084 1.472 0.942 0.897 0.875 0.911 0.945 0.955 0.990 1.022
ARDL 0.862 0.852 0.814 0.917 0.917 0.991 1.031 0.975 1.021 1.096

Third month

Forec. hor. 1 2 3 6 9 12 15 18 21 24

RU-MIDAS 1.060 1.072 0.889** 0.775** 0.944 0.911* 0.816** 0.908** 0.918** 0.905**
MF-VAR-SF 1.092 1.031 0.853 0.755* 0.757 0.823 0.802** 0.860* 0.879* 0.846**
MF-VAR-KF 1.852 1.078 0.921 0.811 0.823 0.887 0.864* 0.942 0.960 0.923
ARDL 0.954 0.850* 0.792** 0.814** 0.833* 0.918 0.890 0.876 0.896 0.908

Note: Boldface indicates improvements over the benchmark AR model. Lag selection is done with the
SIC. Statistically significant reductions in MSPE according to the Diebold-Mariano test are marked using
**(5% significance level) and *(10% significance level). The first estimation sample extends from July 1981
to December 1999, and the evaluation sample extends from January 2000 to November 2013.

34
 Table 6: Forecasting monthly industrial production with the quarterly SPF

All months pooled

Forec. hor. 1 2 3 6 9 12 15 18 21 24

RU-MIDAS 1.035 0.974 0.942* 0.952** 0.935* 0.913 0.928 0.969 0.959 0.966
MF-VAR-SF 1.430 1.141 1.006 0.954 0.912** 0.919** 0.928** 0.939** 0.947** 0.950**
MF-VAR-KF 1.281 1.038 0.905** 0.859** 0.913** 0.940* 0.937** 0.943** 0.949* 0.949*
ARDL 0.931** 0.924* 0.926* 0.928 0.929 0.922 0.928 0.947* 0.953* 0.981

First month

Forec. hor. 1 2 3 6 9 12 15 18 21 24

RU-MIDAS 1.138 1.087 1.085 1.073 0.931** 0.925** 0.944* 0.975 0.953* 0.974**
MF-VAR-SF 1.973 1.314 1.064 1.115 0.961 0.945** 0.951* 0.950* 0.954** 0.966
MF-VAR-KF 1.855 1.128 1.042 0.924 0.925* 0.983 0.972 0.955 0.960 0.965
ARDL 0.932 0.965 1.012 0.990 0.935 0.936 0.943 0.943 0.949* 1.003

Second month

Forec. hor. 1 2 3 6 9 12 15 18 21 24

RU-MIDAS 1.006 0.875** 0.950 0.894** 0.943** 0.922** 0.936* 0.952* 0.964* 0.991
MF-VAR-SF 1.412 1.017 1.066 0.911** 0.914** 0.916** 0.927** 0.939** 0.947** 0.954**
MF-VAR-KF 1.257 1.044 0.892 0.841** 0.924** 0.936** 0.934** 0.940* 0.949 0.947*
ARDL 0.896** 0.899** 0.867* 0.921 0.948 0.936 0.941 0.951 0.960 0.994

Third month

Forec. hor. 1 2 3 6 9 12 15 18 21 24

RU-MIDAS 0.996 0.991 0.803** 0.913** 0.933** 0.895** 0.907* 0.981 0.960 0.932**
MF-VAR-SF 1.151 1.134 0.895* 0.869** 0.868** 0.899** 0.909** 0.927** 0.940** 0.933**
MF-VAR-KF 0.990 0.953 0.793** 0.824** 0.892** 0.907** 0.909** 0.935* 0.938* 0.936*
ARDL 0.949** 0.916* 0.904* 0.887 0.907 0.897 0.903 0.950 0.949 0.947

Note: Boldface indicates improvements over the benchmark AR model. Lag selection is done with the
SIC. Statistically significant reductions in MSPE according to the Diebold-Mariano test are marked using
**(5% significance level) and *(10% significance level). The first estimation sample extends from July 1981
to December 1999, and the evaluation sample extends from January 2000 to November 2013.

35
 Table 7: Forecasting monthly consumption with the quarterly SPF

All months pooled

Forec. hor. 1 2 3 6 9 12 15 18 21 24

RU-MIDAS 1.138 1.053 1.196 0.935* 0.977 1.029 1.019 1.092 1.077 1.014
MF-VAR-SF 1.212 1.214 1.129 0.991 0.979 0.938** 0.951 0.950** 0.945** 0.946**
MF-VAR-KF 1.387 1.291 1.072 1.001 0.956 0.931* 0.947* 0.942** 0.941** 0.945**
ARDL 1.109 1.215 1.226 1.010 1.074 1.080 1.080 1.056 1.069 1.027

First month

Forec. hor. 1 2 3 6 9 12 15 18 21 24

RU-MIDAS 1.087 1.112 1.069 1.027 0.982 1.006 0.993 1.101 1.042 0.999
MF-VAR-SF 0.976 1.213 1.230 0.955 1.007 0.940 0.957* 0.964** 0.961** 0.960**
MF-VAR-KF 1.260 1.141 1.081 0.956 0.987 0.932 0.945 0.958** 0.955** 0.956**
ARDL 0.969 1.480 1.290 1.012 1.051 1.132 1.333 1.068 1.068 1.058

Second month

Forec. hor. 1 2 3 6 9 12 15 18 21 24

RU-MIDAS 1.038 1.161 1.253 0.845** 0.960** 1.061 0.994 1.180 1.166 1.023
MF-VAR-SF 1.212 1.291 1.107 0.970 0.925 0.914** 0.931** 0.959** 0.944** 0.947**
MF-VAR-KF 1.336 1.299 1.135 1.031 0.916 0.914** 0.930** 0.950** 0.941** 0.947**
ARDL 1.217 1.192 1.124 0.981 1.124 1.032 1.052 1.032 1.086 0.998

Third month

Forec. hor. 1 2 3 6 9 12 15 18 21 24

RU-MIDAS 1.241 0.880** 1.343 0.914 0.997 1.020 1.098 0.958* 1.008 1.022
MF-VAR-SF 1.391 1.121 1.001 1.088 1.015 0.971 0.971 0.916** 0.920** 0.925**
MF-VAR-KF 1.510 1.391 0.976 1.040 0.967 0.958 0.974 0.909** 0.917** 0.928**
ARDL 1.160 1.051 1.261 1.048 1.035 1.063 1.033 1.003 1.048 1.023

Note: Boldface indicates improvements over the benchmark AR model. Lag selection is done with the
SIC. Statistically significant reductions in MSPE according to the Diebold-Mariano test are marked using
**(5% significance level) and *(10% significance level). The first estimation sample extends from July 1981
to December 1999, and the evaluation sample extends from January 2000 to November 2013.

36
 Table 8: Forecasting monthly exchange rate with the CPI quarterly SPF

All months pooled

Forec. hor. 1 2 3 6 9 12 15 18 21 24

RU-MIDAS 0.910* 1.037 0.966 0.884 0.940 0.955 0.942 0.979 0.896 0.924
MF-VAR-SF 0.915* 0.964 0.934 0.877* 0.895 0.912 0.887 0.874* 0.878* 0.88*
MF-VAR-KF 1.087 0.964 0.772** 0.859** 0.889** 0.913** 0.919** 0.915** 0.905** 0.902**
ARDL 0.904** 0.944 0.923 0.856* 0.854* 0.908 0.918 0.914 0.924 0.949

First month

Forec. hor. 1 2 3 6 9 12 15 18 21 24

RU-MIDAS 0.829** 1.047 1.043 0.884** 0.954 0.919* 0.915* 0.968 0.903* 0.893**
MF-VAR-SF 0.881 0.985 0.991 0.883 0.887 0.904 0.875 0.878 0.862 0.875
MF-VAR-KF 1.135 1.026 0.858** 0.912** 0.909** 0.921** 0.921** 0.907** 0.909** 0.908**
ARDL 0.835** 0.950 0.946 0.840* 0.854 0.901 0.912 0.947 0.919 0.950

Second month

Forec. hor. 1 2 3 6 9 12 15 18 21 24

RU-MIDAS 0.925** 1.010 0.927 0.854** 0.953 1.054 1.003 1.034 0.930 0.905**
MF-VAR-SF 0.884* 0.961 0.920 0.853** 0.900 0.924 0.916 0.907 0.910 0.906
MF-VAR-KF 0.999 0.904 0.738** 0.827** 0.872** 0.903** 0.916** 0.924** 0.920** 0.916**
ARDL 0.916 0.971 0.936 0.859* 0.858* 0.902 0.924 0.930 0.964 0.966

Third month

Forec. hor. 1 2 3 6 9 12 15 18 21 24

RU-MIDAS 0.997 1.049 0.935** 0.911** 0.915* 0.891** 0.904** 0.930 0.852** 0.973
MF-VAR-SF 1.015 0.940 0.897** 0.892* 0.897 0.906 0.870 0.832 0.859 0.859*
MF-VAR-KF 1.169 0.943 0.727** 0.839** 0.888** 0.914** 0.920** 0.913** 0.885** 0.882**
ARDL 0.981 0.910 0.891* 0.87* 0.849* 0.922 0.919 0.861 0.885 0.931

Note: Boldface indicates improvements over the benchmark RW model. Lag selection is done with the
SIC. Statistically significant reductions in MSPE according to the Diebold-Mariano test are marked using
**(5% significance level) and *(10% significance level). The first estimation sample extends from July 1981
to December 1999, and the evaluation sample extends from January 2000 to November 2013.

37
 Table 9: Forecasting monthly exchange rate with the T-bill quarterly SPF

All months pooled

Forec. hor. 1 2 3 6 9 12 15 18 21 24

RU-MIDAS 0.901** 1.075 1.023 1.022 1.112 1.131 1.197 1.405 1.535 1.728
MF-VAR-SF 0.907* 0.991 0.992 0.982 0.991 0.966 0.930 0.912 0.926 0.947
MF-VAR-KF 1.097 0.931* 0.686** 0.799** 0.846** 0.880** 0.892** 0.888** 0.874** 0.873**
ARDL 0.911** 0.983 1.001 1.038 1.099 1.151 1.164 1.211 1.318 1.417

First month

Forec. hor. 1 2 3 6 9 12 15 18 21 24

RU-MIDAS 0.837** 1.141 1.087 1.031 1.125 1.086 1.184 1.363 1.596 1.718
MF-VAR-SF 0.881 0.997 1.030 0.999 1.008 0.990 0.936 0.922 0.929 0.949
MF-VAR-KF 1.072 0.965 0.773** 0.855** 0.862** 0.878** 0.877** 0.860** 0.861** 0.864**
ARDL 0.841** 0.978 1.022 1.024 1.092 1.128 1.125 1.194 1.304 1.409

Second month

Forec. hor. 1 2 3 6 9 12 15 18 21 24

RU-MIDAS 0.889** 1.087 0.998 1.018 1.135 1.213 1.233 1.491 1.551 1.691
MF-VAR-SF 0.884* 0.986 0.978 0.961 1.004 0.978 0.963 0.949 0.960 0.975
MF-VAR-KF 1.043 0.889 0.655** 0.763** 0.825** 0.870** 0.893* 0.905* 0.902** 0.898**
ARDL 0.914* 1.014 0.996 1.046 1.092 1.147 1.178 1.241 1.355 1.429

Third month

Forec. hor. 1 2 3 6 9 12 15 18 21 24

RU-MIDAS 1.010 0.981 0.989 1.016 1.076 1.092 1.171 1.354 1.457 1.775
MF-VAR-SF 0.983 0.990 0.972 0.984 0.963 0.931 0.888 0.861 0.887 0.914
MF-VAR-KF 1.224 0.929 0.6374** 0.780** 0.852** 0.891** 0.905** 0.898** 0.858** 0.855**
ARDL 1.001 0.960 0.985 1.045 1.114 1.177 1.187 1.195 1.293 1.414

Note: Boldface indicates improvements over the benchmark RW model. Lag selection is done with the
SIC. Statistically significant reductions in MSPE according to the Diebold-Mariano test are marked using
**(5% significance level) and *(10% significance level). The first estimation sample extends from July 1981
to December 1999, and the evaluation sample extends from January 2000 to November 2013.

38
 Table 10: Different aggregation rules for the MF-VAR-KF model

Panel A: High-frequency DGP

(ρ, δh , δl , σ x , σ y ) = (0.5, 0.5, 0.4, 2, 1)

All months First month

Forec. hor. 1 2 3 6 9 12 1 2 3 6 9 12

MF-VAR-KF (MM) 0.881 0.917 0.941 0.971 0.974 0.971 0.503 0.679 0.794 0.904 0.934 0.957
MF-VAR-KF (stock) 0.886 0.943 0.972 1.003 1.008 1.003 0.478 0.680 0.797 0.910 0.945 0.969

Second month Third month

Forec. hor. 1 2 3 6 9 12 1 2 3 6 9 12

MF-VAR-KF (MM) 1.046 1.015 1.000 1.001 1.004 0.979 1.066 1.046 1.007 1.005 0.989 0.989
MF-VAR-KF (stock) 1.005 1.024 1.008 1.011 1.009 1.004 1.162 1.142 1.095 1.069 1.041 1.036

Panel B: Mixed-frequency DGP (δ = 0.2)

All months First month

Forec. hor. 1 2 3 6 9 12 1 2 3 6 9 12

MF-VAR-KF (MM) 0.684 0.675 0.801 0.834 0.878 0.922 0.718 0.556 0.569 0.652 0.790 0.875
MF-VAR-KF (stock) 0.733 0.698 0.814 0.850 0.897 0.929 0.740 0.563 0.600 0.679 0.806 0.884

Second month Third month

Forec. hor. 1 2 3 6 9 12 1 2 3 6 9 12

MF-VAR-KF (MM) 0.639 0.600 0.885 0.896 0.911 0.928 1.076 1.281 0.975 1.032 1.038 1.039
MF-VAR-KF (stock) 0.682 0.631 0.884 0.902 0.914 0.938 0.766 0.895 0.944 0.969 0.965 0.977

Note: This table reports the median of the relative Mean Square Prediction Error for the MF-VAR-KF
models averaged over 1000 replications. The benchmark model is an AR model. The MF-VAR-KF model
is estimated using two different aggregation constraints: (i) Mariano and Murasawa (2003) constraint
(denoted as MF-VAR-KF (MM) model), and (ii) an aggregation rule designed for stock variable (see
equation (31) (denoted as MF-VAR-KF (stock)). Additional details on the DGPs are provided in the text.

39