When are Google data useful to

nowcast GDP? An approach via

pre-selection and shrinkage
Laurent Ferrara1 and Anna Simoni2
April 2019, WP #717
ABSTRACT
Nowcasting GDP growth is extremely useful for policy-makers to assess macroeconomic conditions
in real-time. In this paper, we aim at nowcasting euro area GDP with a large database of Google
search data. Our objective is to check whether this specific type of information can be useful to
increase GDP nowcasting accuracy, and when, once we control for official variables. In this respect,
we estimate shrunk bridge regressions that integrate Google data optimally screened through a
targeting method, and we empirically show that this approach provides some gain in pseudo-real-time
nowcasting of euro area GDP quarterly growth. Especially, we get that Google data bring useful
information for GDP nowcasting for the four first weeks of the quarter when macroeconomic
information is lacking. However, as soon as official data become available, their relative nowcasting
power vanishes. In addition, a true real-time analysis confirms that Google data constitute a reliable
alternative when official data are lacking.

Keywords: Nowcasting, Big data, Sure Independence Screening, Ridge Regularization.

JEL classification: C53, C55, E37

1
Banque de France, [email protected]
2
CREST, CNRS, Ecole Polytechnique, ENSAE, [email protected]
We would like to thank Roberto Golinelli, Michele Lenza, Francesca Monti, Giorgio Primiceri, Simon Sheng,
Hal Varian, Pablo Winant and the participants of the 10th ECB Conference on Macro forecasting with large
datasets, and Data Day@HEC conference for useful comments. We would like to thank Per Nymand-
Andersen (ECB) for sharing the Google dataset as well as Dario Buono and Rosa Ruggeri-Cannata (Eurostat)
for sending the real-time euro area GDP data. We are grateful to Vivien Chbicheb for outstanding research
assistance. Anna Simoni gratefully acknowledges financial support from ANR-11-LABEX-0047 and Fondation
Banque de France for the hospitality. A first version of this paper was circulated under the title:
Macroeconomic nowcasting with big data through the lens of a targeted factor model.

Working Papers reflect the opinions of the authors and do not necessarily express the views of the
Banque de France. This document is available on publications.banque-france.fr/en

Electronic copy available at: https://ssrn.com/abstract=3370917

 NON-TECHNICAL SUMMARY
Figure: Pseudo-real-time analysis with pre-selection of Google data. RMSFEs from models:
(i) with only Google data (green bars), (ii) with all the variables (official and Google data) (red
bars), (iii) with only official variables (blue bars)

Nowcasting GDP growth is extremely useful for policy-makers to assess macroeconomic conditions
in real-time. The concept of macroeconomic nowcasting has been popularized by many researchers
(see e.g. Giannone et al., 2008) and differs from standard forecasting approaches in the sense it aims at
evaluating current macroeconomic conditions on a high-frequency basis.

In the existing literature, GDP nowcasting tools integrate standard official macroeconomic
information stemming, for instance, from National Statistical Institutes, Central Banks, International
Organizations. However, more recently, a lot of emphasis has been put on the possible gain that
forecasters can get from using alternative sources of high-frequency information, referred to as Big
Data (see for example Varian, 2014, Giannone et al., 2017, or Buono et al., 2018). One of the main
sources of alternative data is Google search; seminal papers on the use of such data for forecasting are
the ones by Choi and Varian (2009) and Choi and Varian (2012). Overall, empirical papers show
evidence of some forecasting power for Google data, at least for some specific macroeconomic
variables such as consumption (Choi and Varian, 2012). However, when correctly compared to other
sources of information, the jury is still out on the gain that economists can get from using Google data
for forecasting and nowcasting.

In this paper, we estimate both pseudo real-time and true real-time nowcasts for the euro area
quarterly GDP growth between 2014q1 and 2016q1 by plugging Google data into the analysis, in
addition to official variables on industrial production and opinion surveys, commonly used as
predictors for GDP growth. The approach that we carry out is deliberately extremely simple and relies
on a bridge equation that integrates variables selected from a large set of Google data, as proposed by
Angelini et al. (2011). More precisely, we pre-select Google variables by targeting GDP growth using
the Sure Independence Screening method put forward by Fan and Lv, (2008) enabling to preselect the
Google variables the most related to GDP growth before entering the bridge equation. After pre-
selection we use Ridge regularization to estimate the bridge equation as the number of pre-selected
variables may still be large.

Banque de France Working Paper #717 ii

Electronic copy available at: https://ssrn.com/abstract=3370917

Four main stylized facts come out from our empirical analysis. First, we point out the usefulness of
Google search data for nowcasting euro area GDP for the first four weeks of the quarter when there
is no available official information about the state of the economy. Indeed, we show that at the
beginning of the quarter, Google data provide an accurate picture of the GDP growth rate. Against
this background, this means that such data are a good alternative in the absence of official information
and can be used by policy-makers. Second, we get that as soon as official data become available, that
is starting from the fifth week of the quarter, then the gain from using Google data for GDP
nowcasting rapidly vanishes. This result contributes to the debate on the use of big data for short-term
macroeconomic assessment when controlling for standard usual macroeconomic information. Third,
we show that pre-selecting Google data before entering the nowcasting models appears to be a
pertinent strategy in terms of nowcasting accuracy. Indeed, this approach enables to retain only
Google variables that have some link with the targeted variable. Finally, we carry out a true real-time
analysis by nowcasting euro area GDP growth rate using the official Eurostat timeline and vintages of
data. We show that the three previous results still hold in real-time, in spite of an expected increase in
the size of errors, suggesting that Google search data can be effectively used in practice to help the
decision-making process.

Évaluation en temps réel du PIB avec des

données Google: Une approche par
présélection et réduction de la dimension

RÉSUMÉ
L’évaluation en temps réel (nowcasting) du taux de croissance du PIB est extrêmement utile aux
décideurs de politique économique afin d’appréhender correctement l’activité macroéconomique à une
fréquence élevée. Dans ce travail, nous cherchons à évaluer en temps réel le PIB de la zone euro à
l’aide d’une grande base de données de recherches Google par mots-clés. Notre objectif est de vérifier
si, et quand, ce type d’information permet d’améliorer la précision des évaluations en temps réel
lorsqu’on contrôle par des variables officielles d’enquêtes et de production. Pour cela, nous estimons
des modèles de régression qui permettent une réduction de la dimension à partir des données Google
préalablement présélectionnées de manière optimale, et nous montrons de manière empirique
l’efficacité de cette approche lorsqu’on cherche à évaluer le PIB de la zone euro. En particulier, nous
montrons que les données Google contiennent une information utile pendant les quatre premières
semaines du trimestre, lorsque l’information macroéconomique officielle sur le trimestre en cours
n’est pas disponible. Toutefois, lorsque les données officielles deviennent disponibles, le gain relatif
des données Google se dissipe rapidement. Enfin, nous montrons qu’une analyse dans des vraies
conditions de temps réel, à partir de données millésimées, confirme tous les résultats précédents,
notamment que les données Google constituent une alternative crédible lorsque les données officielles
ne sont pas encore disponibles.

Mots-clés : Évaluation en temps-réel du PIB, Grandes bases de données, Pré-sélection, Réduction de

la dimension

Les Documents de travail reflètent les idées personnelles de leurs auteurs et n'expriment pas nécessairement la
position de la Banque de France. Ils sont disponibles sur publications.banque-france.fr

Banque de France Working Paper #717 iii

Electronic copy available at: https://ssrn.com/abstract=3370917

1 Introduction
Big datasets are now widely used by practitioners for short-term macroeconomic
forecasting and nowcasting purposes. In this paper, we ask the question whether such
data are still useful when controlling for official variables, such as opinion surveys or
production, generally used by forecasters. And if so, when exactly are those alternative
data actually adding a gain in nowcasting accuracy, both in quasi and true real-time
frameworks. In this respect, we focus on Google search data and assess their ability to
provide useful information to nowcast the euro area quarterly GDP growth rate. We
empirically show that they are indeed useful, but only when official data are not avail-
able to practitioners, that is during the first four weeks at the beginning of the quarter.

Nowcasting GDP growth is extremely useful for policy-makers to assess macroe-

conomic conditions in real-time. The concept of macroeconomic nowcasting has been
popularized by many researchers (see e.g. Giannone et al. [2008]) and differs from stan-
dard forecasting approaches in the sense it aims at evaluating current macroeconomic
conditions on a high-frequency basis. The idea is to provide policy-makers with a real-
time evaluation of the state of the economy ahead of the release of official Quarterly
National Accounts, that always come out with a delay. For example, the New York Fed
and the Atlanta Fed have recently developed new tools in order to evaluate US GDP
quarterly growth on a high-frequency basis1 . The tool developed by the Atlanta Fed,
referred to as GDPNow, is updated 6 to 7 times per month, while the NY Fed’s tool is
updated every Friday. With reference to countries other than the US, many papers have
put forward econometric modelling to nowcast GDP growth in advanced countries (see
among others Frale et al. [2010] or Kuzin et al. [2011] for the euro area, Aastveit and
Trovik [2012] for Norway or Bragoli [2017] for Japan), as well as in emerging countries
(see for example Modugno et al. [2016] for Turkey or Bragoli et al. [2015] for Brazil).
Some researchers have also proposed approaches to nowcast economic output at a global
level in order to assess on a regular high-frequency basis world economic conditions (see
Ferrara and Marsilli [2018] or Golinelli and Parigi [2014]).
In the existing literature, GDP nowcasting tools integrate standard official macroeco-
nomic information stemming, for instance, from National Statistical Institutes, Central
Banks, International Organizations. Typically, three various sources of official data
are considered: (i) hard data, like production, sales, employment, (ii) opinion surveys
(households or companies are asked about their view on current and future economic
conditions), and (iii) financial markets information (sometimes available on high fre-
1
See the websites https://www.newyorkfed.org/research/policy/nowcast and https://www.
frbatlanta.org/cqer/research/gdpnow.aspx

Electronic copy available at: https://ssrn.com/abstract=3370917

quency basis). However, more recently, a lot of emphasis has been put on the possible
gain that forecasters can get from using alternative sources of high-frequency informa-
tion, referred to as Big Data (see for example Varian [2014], Giannone et al. [2017] or
Buono et al. [2018]). Various sources of Big Data have been used in the recent literature
such as for example web scraped data, scanner data or satellite data. One of the main
source of alternative data is Google search and seminal papers on the use of such data
for forecasting are the ones by Choi and Varian [2009] and Choi and Varian [2012] (see
also Scott and Varian [2015] who combine Kalman filters, spike-and-slab regression and
model averaging to improve short-term forecasts).
Overall, empirical papers show evidence of some forecasting power for Google data,
at least for some specific macroeconomic variables such as consumption (Choi and Var-
ian [2012]), unemployment rate (D’Amuri and Marcucci [2017]), building permits (Coble
and Pincheira [2017]) or car sales (Nymand-Andersen and Pantelidis [2018]). However,
when correctly compared with other sources of information, the jury is still out on the
gain that economists can get from using Google data for forecasting and nowcasting. For
example, Vosen and Schmidt [2011] show that Google Trends data lead to an accuracy
gain when compared with business surveys to forecast the annual growth rate of US
household consumption. But some other papers tend to show that the gain in forecast-
ing using Google data is very weak when other sources of information are accounted for
in the analysis. For example, Goetz and Knetsch [2019] estimate German GDP using
simultaneously both official and Google data on a monthly basis and show that adding
Google data only leads to limited accuracy gains. However, they provide some evidence
that those data can be a potential alternative to survey variables. We also refer to Li
[2016] on this issue. Overall, it seems that Google data can be extremely useful when
economist do not have access to information or when information is fragmented, as for
example when dealing with emerging economies (see Carriere-Swallow and Labbe [2013])
or low-income developing countries (Narita and Yin [2018]).

In this paper, we estimate both pseudo real-time and true real-time nowcasts for the
euro area quarterly GDP growth between 2014q1 and 2016q1 by plugging Google data
into the analysis, in addition to official variables on industrial production and opin-
ion surveys, commonly used as predictors for GDP growth. Google data are indexes
of weekly volume changes of Google searches by keywords in the six main euro area
countries about different topics which are gathered in 26 broad categories such as auto
and vehicles, finance, food and drinks, real estate, etc. Those broad categories are then
split into a total of 269 sub-categories per country, leading to a total of 1776 variables

Electronic copy available at: https://ssrn.com/abstract=3370917

for each country2 . Our objective is to assess whether Google search data bring some
gain in nowcasting accuracy and when. The approach that we carry out is deliberately
extremely simple and relies on a bridge equation that integrates variables selected from
a large set of Google data, as proposed by Angelini et al. [2011]. More precisely, we pre-
select Google variables by targeting GDP growth in the vein of Bai and Ng [2008] but
with a different approach. Pre-selection is implemented by using the Sure Independence
Screening method put forward by Fan and Lv [2008] enabling to preselect the Google
variables the most related to GDP growth before entering the bridge equation. After
pre-selection we use Ridge regularization to estimate the bridge equation as the number
of pre-selected variables may still be large.

Four main stylized facts come out from our empirical analysis. First, we point out
the usefulness of Google search data for nowcasting euro area GDP for the first four
weeks of the quarter when there is no available official information about the state of
the economy. Indeed, we show that at the beginning of the quarter, Google data provide
an accurate picture of the GDP growth rate. Against this background, this means that
such data are a good alternative in the absence of official information and can be used
by policy-makers. Second, we get that as soon as official data become available, that
is starting from the fifth week of the quarter, then the gain from using Google data for
GDP nowcasting rapidly vanishes. This result contributes to the debate on the use of
big data for short-term macroeconomic assessment when controlling for standard usual
macroeconomic information. Third, we show that pre-selecting Google data before en-
tering the nowcasting models appears to be a pertinent strategy in terms of nowcasting
accuracy. Indeed, this approach enables to retain only Google variables that have some
link with the targeted variable. This result confirms previous analyses that have been
carried when dealing with large datasets through dynamic factor models (see e.g. Bai
and Ng [2008] or Schumacher [2010]). Finally, we carry out a true real-time analysis by
nowcasting euro area GDP growth rate using the official Eurostat timeline and vintages
of data. We show that the three previous results still hold in real-time, in spite of an
expected increase in the size of errors, suggesting that Google search data can be effec-
tively used in practice to help the decision-making process.

The rest of the paper is organized as follows. In Section 2 we describe the model we
consider for nowcasting, the Sure Independence Screening (SIS) approach to pre-select
the data, as well as the Ridge regularization. Section 3 describes the structure of the

2
See for example Bontempi et al. [2018] for a detailed description of this dataset, in a different
framework

Electronic copy available at: https://ssrn.com/abstract=3370917

Google search data used for nowcasting. The empirical results are presented in Section
4 and Section 5 concludes.

2 Methodology
2.1 The nowcasting approach
In order to get GDP nowcasts, we focus on linear bridge equations that link quarterly
GDP growth rates and monthly economic variables. The classical bridging approach is
based on linear regressions of quarterly GDP growth on a small set of key monthly in-
dicators as for example in Diron [2008]. In our exercise, in addition to those monthly
variables, we also consider Google data, available at a higher frequency, and we aim at
assessing their nowcasting power. More precisely, Google data are available on a weekly
basis, providing thus additional information when official information is not yet avail-
able. Even if Google data are not on average extremely correlated with the GDP growth
rate, we are going to show that they still provide accurate GDP nowcasts if conveniently
treated.

Therefore, we assume that we have three types of data at disposal: soft data, such as
opinion surveys, hard data, such as industrial production or sales, and data stemming
from Google search machines. Let t denote a given quarter of interest identified by its
last month, for example the first quarter of 2005 is dated by t = March2005. A general
model to nowcast the growth rate of any macroeconomic series of interest Yt for a specific
quarter t is the following, for t = 1, . . . , T :

Yt = β0 + βs0 xt,s + βh0 xt,h + βg0 xt,g + εt , E[εt |xt,s , xt,h , xt,g ] = 0, (2.1)

where xt,s is the Ns -vector containing soft variables, xt,h is the Nh -vector containing
hard variables, xt,g is the Ng -vector of variables coming from Google search and εt is an
unobservable shock. In our empirical analysis Yt is the quarterly GDP growth rate of
the euro area. Because variables xt,s , xt,h and xt,g are sampled over different frequencies
(monthly vs weekly), the relevant dataset for calculating the nowcast evolves within
(w)
the quarter. By denoting with xt,j , j ∈ {s, h, g}, the j-th series released at week
w = 1, . . . , 13 of quarter t, we denote the relevant information set at week w of a quarter
t by
(w) (w)
Ωt := {xt,j , j ∈ {s, h, g} such that xt,j is released at w}.
(w)
For simplicity, we keep in Ωt only the observations relative to the current quarter t

Electronic copy available at: https://ssrn.com/abstract=3370917

 (w)
and do not consider past observations. While xt,g is in Ωt for every w = 1, . . . , 13, the
other variables are in the relevant information set only for the weeks corresponding to
their release and so the dataset is unbalanced.

To explicitly account for the different frequencies of the variables, we replace model
(2.1) by a model for each week w such that:

(w)
Ybt|w = E[Yt |Ωt ], t = 1, . . . , T and w = 1, . . . , 13
(w) 0 (w)
0 0 (w) (w)
and E[Yt |Ωt ] = β0,w + βs,w xt,s + βh,w xt,h + βg,w xt,g (2.2)

(w) (w) (w)

where xt,j = 0 if xt,j ∈ / Ωt . For instance, as the first observation of industrial pro-
(w)
duction relative to the current quarter t is only released in week 9, then we set xt,h = 0
for every w = 1, . . . , 8. The bridge equation (2.2) exploits weekly information to obtain
more accurate nowcasts of quarterly GDP growth.

2.2 Pre-selection of Google data

The recent literature on nowcasting and forecasting with large datasets comes to the
conclusion that using the largest available dataset is not necessarily the optimal approach
when aiming at nowcasting a specific macroeconomics variable such as GDP, at least in
terms of nowcasting accuracy. For example, against the background of bridge equations
augmented with dynamic factors, Barhoumi et al. [2010] empirically show that factors
estimated on a small database lead to competitive results for nowcasting French GDP
compared with the most disaggregated data. From a theoretical point of view, Boivin
and Ng [2006] suggest that larger databases lead to poor forecast when idiosyncratic
errors are cross-correlated or when the forecasting power comes from a factor that is
dominant in a small database but is dominated in a larger dataset. An empirical way
to circumvent this issue is to target more accurately the variable to be nowcast. For
example, Bai and Ng [2008] show that forming targeted predictors enables to improve
the accuracy of inflation forecasts while Schumacher [2010] shows that targeting German
GDP within a dynamic factor model is a performing strategy.
In this respect, all the categories and subcategories in the Google search data are not
necessarily correlated with the GDP growth that we want to nowcast. Therefore, using
all the variables in the Google search dataset is not necessarily a good strategy because
one would pay the price of dealing with ultra-high dimensionality without increasing the
nowcasting accuracy as measured by the Mean Squared Forecasting Error (MSFE). For
this reason we consider a pre-selection procedure before using data for nowcasting, that

Electronic copy available at: https://ssrn.com/abstract=3370917

is, a procedure enabling to select a subset of the variables in the Google search dataset
that are the most relevant for GDP growth nowcasting. In a second step, we will use
a Ridge regularization to estimate model (2.2) by using the selected subset of Google
data, which is the most “related” with the variable Yt and which captures much of the
variability in GDP growth. As explained in Section 2.3 below, a regularization technique
is required because the number of selected variables can still be large while not ultra-high.

While in our empirical analysis we have tried several pre-selection procedures, it

turns out that the innovative approach put forward by Fan and Lv [2008] appears to
provide interesting and intuitive results. This approach is referred to as Sure Inde-
pendence Screening, or SIS hereafter. Sure screening refers to the property that all
important variables survive after applying a variable screening procedure with probability
tending to 1 (see Fan and Lv [2008], p. 853). The basic idea of this approach is based
on correlation learning and relies on the fact that only the variables with the highest
absolute correlation should be used in modelling.

Let us start from a standard linear regression equation with only the standardized
Ng Google variables as explanatory variables, that is β0 = βs = βh = 0 in equation
(2.1). Let M ∗ = {1 ≤ j ≤ Ng : βg,j 6= 0} be the true sparse model with non-sparsity size
s = |M ∗ |. The other Ng − s variables can also be correlated with Y via linkage to the
predictors contained in the true sparse model. Let Y denote the T -vector of quarterly
GDP growth: Y = (Y1 , . . . , YT )0 . We compute ω = (ω1 , . . . , ωNg )0 , the vector of marginal
correlations of predictors with the response variable Yt , such as

0
ω = X g Y, (2.3)

where X g is the T × Ng matrix of average Google data where the average is taken over
each quarter and that then has been centered and standardized columnwise. The average
over each quarter is taken to make the weekly Google data comparable to the quarterly
GDP growth data in terms of frequency. For any given λ ∈]0, 1[, the Ng componentwise
magnitudes of the vector ω are sorted in a decreasing order and we define a submodel
Mλ such as: Mλ = {1 ≤ j ≤ Ng : |ωj | is among the first [λT ] largest of all }, where [λT ]
denotes the integer part of λT . Since only the order of componentwise magnitudes of
ω is used, this procedure is invariant under scaling and thus it is identical to selecting
predictors using their correlations with the response. This approach is an easy way to
filter out Google variables with the weaker correlations with GDP growth rate so that we
are left with d = [λT ] < T Google variables. An important feature of the SIS procedure
is that it uses each covariate xt,g,j independently as a predictor to decide how useful it

Electronic copy available at: https://ssrn.com/abstract=3370917

is for predicting Yt .
This method is desirable because it has the sure screening property, that is, with
probability tending to one, all the important variables in the true model are retained
after applying this method. In fact, Fan and Lv [2008, Theorem 1] show that under
Normality of εt and other conditions (see Fan and Lv [2008, Conditions 1-4]) the sure
screening property holds, namely for a given λ:

P (M ∗ ⊂ Mλ ) → 1

as Ng → ∞. In particular, SIS can reduce the dimension to [λT ] = O(T 1−θ ) < T for
some θ > 0 and the reduced model Mλ still contains all the variables in the true model
M ∗ with a probability converging to one.
In the following, we write Xt,Mλ = [1, x0t,s , x0t,h , x0t,g,Mλ ]0 , where xt,g,Mλ = {xt,g,j ; j ∈
Mλ } is the vector containing only the selected Google variables. Moreover, for a vector
β ∈ Rp and a set M ⊂ {1, . . . , p} we write M c for the complement of M in {1, . . . , p}
and βM = {βj ; j ∈ M }. The empirical choice of the hyperparameter λ is discussed in
subsection 3.3.

2.3 Ridge regression

Google search data have an extremely large dimension, with the number of variables
much larger than the number of observations (i.e. Ng >> T , sometimes referred to
as fat datasets). Therefore, when using Google search data for nowcasting one has to
deal with such high dimensionality. Even after implementing the pre-selection described
in subsection 2.2, the number of Google variables remains large compared to the time
dimension T . Therefore, one needs to use a machine learning technique suitable to treat
fat datasets.
One of the most popular ways to deal with a large number of covariates and possibly
problems of multicollinearity is the Ridge regression. Let β = (β0 , βs0 , βh0 , βg0 )0 . Ridge
regression estimates equation (2.1) by minimizing a penalized residuals sum of squares
where the penalty is given by the Euclidean squared norm k · k:
( T
)
1X
2
βb = arg min Yt − β0 − βs0 xt,s − βh0 xt,h − βg0 xt,g + αkβk2 ,
β T t=1

where α > 0 is a regularization parameter that tunes the amount of shrinkage. The
estimated coefficients in βb are then shrunk towards zero. By using model (2.2) for each

Electronic copy available at: https://ssrn.com/abstract=3370917

week w and the pre-selection methodology in Section 2.2 we can compute the Ridge
estimator after model selection:
( T 
)
(w) 1 X (w) (w) (w)
2
βbRidge := arg min Yt − β0 − βs0 xt,s − βh0 xt,h − βg0 xt,g + αkβk2 .
β;βg,j =0,j∈Mλc T t=1

(w) (w)0 (w)0

Equivalently, we can write βbRidge = (βbRidge,Mλ , βbRidge,M c )0 where
λ

T
!−1 T
(w) 1X 0 1X 0 (w)
βbRidge,Mλ = Xt,Mλ Xt,M + αI X Yt , βbRidge,M c = 0
T t=1 λ
T t=1 t,Mλ λ

and I is the |Mλ |-dimensional identity matrix. This is the estimator we are going to use
in our empirical analysis. Even if it depends on α in a crucial way, we leave implicit this
dependence. The empirical choice of the hyperparameter α is a crucial issue because it
has an important impact on the nowcasting accuracy. We discuss this choice in Section
3.3

3 Design of the empirical analysis

This section first describes the data used in the empirical analysis. Then, it describes
how to deal with the various reporting lags. Finally, we propose a way to select both
hyperparameters λ and α involved in the estimation procedure.

3.1 Data
Our objective in this paper is to assess the role of Google data for nowcasting the
euro area GDP, especially to assess (i) if these big data are relevant when there is no
official data available for the forecaster and (ii) to what extent these data provide use-
ful information when official data become available. In this respect, the variable Yt in
model (2.1)-(2.2) that we target is the quarterly growth rate of the real euro area GDP,
stemming from Eurostat. The official data that we consider are of two kinds: industrial
production for the euro area as a whole provided by Eurostat, which is a global measure
of hard data and is denoted by IPt , and a composite index of opinion surveys from var-
ious sectors computed by the European Commission (the so called euro area Sentiment
Index) denoted by St .

Electronic copy available at: https://ssrn.com/abstract=3370917

 Our big dataset covers Google searches for the six main euro area countries: Bel-
gium, France, Germany, Italy, Netherlands and Spain. We have at disposal a total of
Ng = 1776 variables, corresponding to 26 categories and 296 subcategories for each coun-
try. Google search data are data related to queries performed with Google search. The
data are indexes of weekly volume changes of Googles queries grouped by category and
by country. Data are normalized at 1 at the first week of January 2004 which is the first
week of availability of these data. Then, the following values indicate the deviation from
the first value. However, there is no information about the search volume. Google data
are weekly data that are received and made available by the European Central Bank
every Tuesday. Original data are not seasonally adjusted, thus we take the growth rate
over 52 weeks to eliminate the seasonality within the data.3
We use data from 20 March 2005 (twelfth week of the first quarter) until 29 March
2016 (thirteen week of the first quarter). We split the sample in two parts and use data
starting from the first week of January 2014 for the out-of-sample analysis.

3.2 Dealing with various reporting lags

An important feature of all these data is that they are released with various report-
ing lags, leading thus to non-balanced information dataset at each point in time. In the
literature, this issue is refereed to as ragged-edge database (see Angelini et al. [2011]).
For instance, Google search data are weekly data available every Tuesday, while the soft
and hard data are monthly data available at the end of every month and at the middle
of the third month of the quarter, respectively. Treating weekly data is particularly
challenging as the number of entire weeks present in every quarter is not always the
same, and a careful analysis has to be done when incorporating these data. In addi-
tion, there is a frequency mismatch in the data as the explained variable is quarterly
and the explanatory variables are either weekly (Google data) or monthly (hard and
soft variables). In order to account for the various frequencies and the timing at which
the predictive variables become available, we adopt the strategy to consider a different
model for every week of the quarter as described in Section 2.1. Thus we end up with
thirteen models given in equation (2.2), each model including the variables available at
this date.

As regards the dates of availability, we mimic the exact release dates as published by

3
Applying this standard seasonal filter eliminates a large part of seasonal effects in the dataset. We
also test for outliers in our study, but dropping detected outliers does not seem to improve nowcasting
accuracy. Obviously this question needs to be tackle in more details in further research.

Electronic copy available at: https://ssrn.com/abstract=3370917

Eurostat. This means that the first survey of the quarter, referring to the first month,
typically arrives in week 5. Then, the second survey of the quarter, related to the second
month, is available in week 9. Industrial production for the first month of the quarter
is only available about 45 days after the end of the reference month, that is generally in
week 11. Finally, the last survey, related to the third month of the quarter, is available
in week 13. A scheme of the release timeline is presented in Figure 1.

By denoting with xt,g,Mλ ,w the vector of pre-selected variables from the Google search
(w)
data for week w of quarter t we construct the variable xt,g in equation (2.2) as the aver-
(w) P
age of the vector of selected Google variables up to the w-th week: xt,g = v≤w xt,g,Mλ ,v .
(3)
That is, take for instance w = 3 (i.e. Model 3 which is used at week 3), then xt,g is
equal to (xt,g,Mλ ,1 + xt,g,Mλ ,2 + xt,g,Mλ ,3 )/3.4 The other variables in equation (2.2) denote,
(w)
respectively: Yt the euro area GDP growth rate, xt,s the monthly data from surveys,
(w)
available at the end of each month, and xt,h denotes the growth rate of the index of
industrial production, available about 45 days after the end of the reference month. Be-
cause of the frequency mismatch within the whole dataset, the thirteen models include a
(w)
different number of predictors, as we have explained above. As regards the survey, xt,s ,
(w)
and the industrial production, xt,h , we impose the following specific structure which
mimics the data release explained above, and that will be used throughout our exercise.
(w)
The variable xt,s is not present in models 1 to 4 because the survey is not available in the
(1) (2) (3) (4)
first four weeks of the quarter, so that βt,s = βt,s = βt,s = βt,s = 0. Then, for models
(w) (w)
5 to 8, xt,s is the value of the survey for the first month of the quarter: xt,s = St,1

Figure 1: Timeline of data release in the pseudo real-time exercise within the quarter.

4
In our empirical analysis, we also test models that do not use the average over weeks of Google
search data as explanatory variables, but instead, Google search data for each new weeks is considered
as the variable for the quarter. Results clearly point that models integrating averaged Google search
data give smaller Mean Squared Forecasting Errors than models that do not use the averaged Google
search data.

10

Electronic copy available at: https://ssrn.com/abstract=3370917

where St,i denotes the variable St referring to the i-th month of quarter t. In models 9
(w)
to 12, xt,s will be equal to the average of the survey data available at the end of the
(w) (w)
first and second month of the quarter: xt,s = (St,1 + St,2 )/2. Last, in model 13, xt,s is
(w)
the average of the survey data over the quarter: xt,s = (St,1 + St,2 + St,3 )/3. Similarly,
(w) (1) (10)
the variable xt,h is not present in models 1 to 10 (so that βt,h = . . . = βt,h = 0) and
(w)
in models 11 to 13, xt,h will be the value of the growth rate of the index of industrial
production IPt for the first month of the quarter.
The idea of having thirteen models is that a researcher will use one of these thirteen
models to nowcast the current-quarter values of Yt depending on the current week of the
quarter. For instance, to nowcast the current-quarter value of Yt at the end of week 2,
the Model w = 2 will be used. In Table 1 in the Annex, we give the thirteen models
based on equation (2.2) and we denote them by M 1, . . . , M 13.

One of the main issue in the literature on big data is to know whether and when
such alternative data are able to bring an additional gain with respect to standard types
of variables, like hard and soft data. To contribute to the existing literature on this is-
sue, we have also estimated nowcasting models without including the vector of variables
selected from the Google search data. That is, these models only include as predictors
(w)
the survey and the growth rate of the index of industrial production (i.e. βt,g = 0 in
equation (2.2)). We have in total four such models, one for each release of data of these
two variables within the quarter, denoted N oGoogle1 , . . . , N oGoogle4 in Table 2 in the
Annex, that will be used for comparison purposes.

An additional issue with the reporting lags concerns the release of GDP figures. In
fact, the first GDP assessment is generally released about 45 days after the end of the
reference quarter, but sometimes the delay may be longer. For instance, GDP figures
for the first quarter of 2014 were only released on the 4th of June 2014. For this reason
if one wants to nowcast in real-time GDP growth for 2014q2 it is not possible to use
the fitting computed with the data available up to 2014q1 because one does not observe
the GDP for 2014q1. Instead, one has to use the estimated parameters computed with
the data available up to 2013q4. Because of this, we impose a gap of two quarters be-
tween the sample used for fitting the model (training sample) and the sample used for
the out-of-sample analysis (test data). For coherence, we use this structure in both the
pseudo-real-time and the true real-time analysis.

Another issue concerns the inclusion of lagged GDP among the explanatory vari-
ables. Because of the delay in the release of the GDP we cannot include the lagged GDP

11

Electronic copy available at: https://ssrn.com/abstract=3370917

as explanatory variable in every nowcasting model. In addition to this, the GDP is not
released at a fixed date, meaning that the release is different at every period (every
quarter and every year). For these reasons we have not included the lagged GDP among
the explanatory variables in the thirteen models (2.2) for the pseudo-real-time analysis.
On the other hand, for the true real-time analysis we have exploited the additional infor-
mation arising from lagged GDP and have included it among the explanatory variables
when it is available. We provide in Figure 3 in the Annex an overview of the dates at
which specific GDP figures are released as well as the indication of the time from which
we can include the lagged GDP among the explanatory variables and the arrival times
of new vintages. We have used this calendar to construct our real-time analysis.
In fact we carry out two types of real-time analysis: (I) a true real-time analysis
which includes the lagged GDP growth when it is available, and (II) a true real-time
analysis which does not include the lagged GDP growth. The latter is meant for compar-
ison with the pseudo-real-time analysis which does neither include lagged GDP values.

3.3 Selection of the tuning parameters λ and α

To construct our estimator of the thirteen models (2.2) two tuning parameters, or
hyperparameters, have to be fixed: λ and α. We select them by using a data-driven
method based on a grid-search procedure over the last training period. We select a pair
of values for (λ, α) for each model and for each nowcasting period. Hence, in total we
have 13 ∗ 9 = 117 values for the pair (λ, α).

The empirical choice of the hyperparameter α is a crucial issue because it has an

important impact on the nowcasting accuracy. Ideally, we would like to choose a value
for α for which the MSFE is as small as possible. Therefore, we follow Li [1986, 1987]
and use the Generalized cross-validation (GCV) technique to select α. This technique
has recently been used by Carrasco and Rossi [2016] in a forecasting setting. The idea
is to select the value of α that minimizes the following quantity:
(w)
T −1 Tt=1 (Yt − Xt0 βbRidge )2
P
b(w) (α)
Q = 
T 2
1 − T −1 tr(R
bT (α))

where tr(·) denotes the trace operator and R

bT (α) is given by

T
!−1 T
X X
R
bT (α) = XM
λ
T −1 0
Xt,Mλ Xt,M λ
+ αI T −1 0
Xt,M λ
t=1 t=1

12

Electronic copy available at: https://ssrn.com/abstract=3370917

 b(w) (α) over a grid of 31
and XMλ = [X1,Mλ , . . . , XT,Mλ ]. In our analysis we minimize QT
(w) b(w) (α).
equispaced values in [0.09, 1.1] denoted by A, so that αbT = arg minα∈A Q T

For λ we consider a grid of 99 equispaced values in (0, 1], denoted by Λ. The selec-
tion is made sequentially: for each value of λ in the grid we select for α in model w the
(w) (w) b(w) (α). This is done for each of the thirteen
value αbT that solves α bT := arg minα∈A Q T
models and for each nowcasting period by using only the training sample corresponding
to the specific nowcasting period we are considering. We notice that T depends on the
nowcasting period.
(w)
Once a value α bT is selected for each value of λ in the grid, we select the value
of λ that minimizes the MSFE for the GDP growth of the last quarter of the training
(w)
sample obtained by using the selected α bT . That is, if T denotes the last quarter of
the training sample, then we select for λ in model w the value λ b(w) = arg minλ∈Λ (YT −
T
0 (w) (w) 2
XT,Mλ βRidge,Mλ (b
b αT )) .

4 Empirical Results
In this section we present the results of our empirical exercises aiming at nowcasting
the euro area GDP growth using various types of data sources. This section is split into
three parts. First, we look at the accuracy gains stemming from using Google data when
controlling for standard official macroeconomic data, by comparing nowcasts obtained
with and without such data, in a pseudo real-time exercise. Then we look at the effects
of pre-selecting Google data before estimating Ridge regressions. Third, we perform a
true real-time analysis.

4.1 Is there a gain from using Google data, and when ?

In this subsection we compare the evolution over the quarter of weekly Root MS-
FEs (RMSFE) stemming from the nowcasting models, with and without Google search
data. We do this exercise in pseudo-real time, that is, by using historical data but
by accounting for their ragged-edge nature. To evaluate the impact of Google search
data on current-quarter nowcasts of the GDP growth, we make two types of compar-
isons. First, we estimate the thirteen nowcasting models by using only Google data,
(w) (w)
that is, xt,s = xt,h = 0 for every w = 1, . . . , 13 in Equation (2.2). Second, to assess
the marginal gain of integrating Google data, we compare the four models that only ac-

13

Electronic copy available at: https://ssrn.com/abstract=3370917

count for hard and soft data (i.e. without Google data) with the corresponding models
given by (2.2) accounting for the full set of information (Google, Survey and Indus-
trial Production). More precisely, we directly compare the following pairs of models:
(N oGoogle1 , M 5), (N oGoogle2 , M 9), (N oGoogle3 , M 11), and (N oGoogle4 , M 13). The
results of these comparisons are reported in Figure 2 below and in Panel 1 of Figure 8 in
the Annex. The estimation has been conducted by using Ridge regularization coupled
with the SIS pre-selection approach as described in Sections 2.2 and 2.3.

Figure 2: The importance of Google data. Pseudo-real-time analysis with pre-selection of Google
data. RMSFEs from: (i) models M1 - M13 with only variables extracted from Google data (in light
gray), (ii) models M1 - M13 with all the variables (St , IPt and Google data) (in gray), (iii) models with
only official variables N oGoogle1 - N oGoogle4 (in black).

The first striking feature that we observe in Figure 2 is the downward sloping evolution
of RMSFEs stemming from the models with full information (Google, Industrial Pro-
duction and Survey) over the quarter. This is in line with what could be expected from
nowcasting exercises when integrating more and more information throughout the quar-
ter (see Angelini et al. [2011]). When using Google information only (light gray bars),
we still observe a decline but to a much lower extent and the RMSFEs stay above 0.25
even at the end of the quarter. However, when focusing on the beginning of the quarter,
models that only integrate Google information provide reasonable RMSFEs that do not
exceed 0.30 (see Figure 2). This result shows that Google search data possess a informa-
tional content that can be valuable for nowcasting GDP growth for the first four weeks

14

Electronic copy available at: https://ssrn.com/abstract=3370917

of the quarter, when there is no other available official information about the current
state of the economy. However, it is striking to see that when information about the
first survey of the quarter arrives, that is in week 5, the model that only incorporates
Google data clearly suddenly underperfoms. Looking at Table 8, we see that the RMSFE
goes from 0.2887 in week 4 when only Google information is used to 0.2361 in week 5
when the full information model is used. In addition, we note that a simple model, only
accounting for official hard and soft information, leads to a much lower RMSFE in week
5 (equal to 0.1807, see Panel 1 of Figure 8 in the Annex). Comparing black bars and
gray bars in Figure 2 clearly shows evidence that there is no additional gain of adding
Google data to the model starting from week 5; a simple model with only hard and soft
information cannot indeed be outperformed.

4.2 Is it worth to pre-select Google data?

As mentioned in Section 2.2, the literature suggests that it could be useful to first
pre-select a sub-sample of Google data before estimating the thirteen models (2.2). In
this respect, various approaches have been put forward in order to target ex ante the
variable of interest (see e.g. Bai and Ng [2008] or Schumacher [2010], against the back-
ground of bridge equations augmented with dynamic factors). In this section we present
the performance of a Ridge regularization approach for nowcasting GDP growth when it
is coupled with the SIS pre-selection method of Fan and Lv [2008] described in Section
2.2, compared with a standard Ridge regularization approach without any pre-selection.

The idea of the SIS pre-selection method is to identify ex ante specific Google vari-
ables, among the initial large dataset, that have the highest absolute correlation with
the targeted variable, namely the GDP growth rate. First, let us have a look at the
relationship between the number of selected variables through the SIS procedure and
the absolute correlation between each Google variable and the GDP growth rate at the
same quarter. We recall that for the Google variables we take the average over each
quarter, see Section 2.2. This relationship is described in Figure 3. We clearly observe
an inverse non-linear relationship, with a kind of plateau starting from an absolute cor-
relation of about 0.25. Indeed, most of Google variables present an absolute correlation
with current GDP growth rate lower than 0.30. Thus it seems useful to only focus on a
core dataset with the highest correlations.

We then analyse the performances in terms of RMSFEs from bridge regressions that use
the SIS pre-selection approach associated with Ridge regularization. Figure 4 presents

15

Electronic copy available at: https://ssrn.com/abstract=3370917

Figure 3: Plot of the number of selected Goggle variables by the SIS method versus the correlation
with current GDP (computed for the first training sample).

the evolution over the 13 weeks of the quarter of RMSFEs stemming from bridge models
estimated using Google search data and Ridge regression coupled or not with the SIS
pre-selection approach. We clearly see that the SIS pre-selection approach (gray bars,
similar to the gray bars in Figure 2) allows for an overall improvement in nowcasting
accuracy. A striking result is that the RMSFE is lower for all the weeks when the SIS
pre-selection approach is used. Moreover, when pre-selection is implemented, RMSFEs
evolve over the quarter in a more smoother way. For example, without any pre-selection,
we observe that in week 6 the RMSFE jumps to 0.3829, from 0.3239 in week 5. The
overall gain underlines the need for pre-selecting data using a targeted approach. Panel
2 in Figure 8 in the Annex reports the exact values of the RMSFEs with and without
pre-selection.

4.3 A true real-time analysis

In this subsection, we carry a true real-time analysis by using vintages of data for
GDP and industrial production5 and by accounting for the observed timeline of data
release as provided by Eurostat. As regards the dates of the GDP release, there is a
large heterogeneity from one period to the other. When available, we also include the
lagged GDP growth among the explanatory variables of the nowcasting models. Figure

5
Survey data are generally not revised.

16

Electronic copy available at: https://ssrn.com/abstract=3370917

Figure 4: Pseudo-real-time: is it worth to preselect? Evolution over the 13 weeks of the quarter
of the RMSFEs stemming from bridge models using Google data, St and IPt estimated from Ridge
regularization with and without SIS pre-selection approaches.

3 in the Annex gives the exact weeks in the out-of-sample period 2014q1-2016q1 where
the lagged GDP growth is included in the real-time analysis.

In Figure 5 we show that pre-selecting Google data is still worth in real-time. Indeed,
RMSFEs obtained from models integrating pre-selected Google data are systematically
lower, for all weeks, than those obtained without any pre-selection. The corresponding
RMSFE values are reported in Panel 3 of Table 8.

In Figure 6, we show the impact of Google search data on GDP growth nowcast-
ing accuracy in the context of a true real-time nowcasting analysis. The corresponding
RMSFE values are reported in Panel 4 of Table 8. Similarly to the pseudo real-time
exercise, we get that during the first 4 weeks of the quarters, when only Google infor-
mation is available, RMSFEs are quite reasonable. This fact is reassuring about the
real-time use of Google search data when nowcasting GDP. However, starting from week
5, as soon as the first survey of quarter is released, the marginal gain of using Google
data instantaneously vanishes. 6

6
There is an exception in week 11, where it is surprising to note that the integration of surveys, past
GDP value and industrial production tend to suddenly increase the RMSFEs, in opposition to what
can be expected from previous empirical results. This stylized has to be further explored.

17

Electronic copy available at: https://ssrn.com/abstract=3370917

Figure 5: True real-time analysis: is it worth to preselect? Evolution over the 13 weeks of the
quarter of the RMSFEs stemming from bridge models using Google data, St , IPt , and lagged GDP
growth estimated from Ridge regularization with and without SIS pre-selection approaches. The models
include the lagged GDP growth when it is available.

Finally, in order to compare the results of the real-time analysis with the ones from
the pseudo-real-time analysis, we compute GDP growth nowcasts without including the
lagged GDP growth among the explanatory variables. The results are given in Figure
7. The corresponding RMSFE values are reported in Panel 5 of Table 8. We see that
both analyses lead to a similar shape in the evolution of RMSFEs within the quarter,
although, as expected, the uncertainty around weekly nowcasts is a bit higher in real
time.

18

Electronic copy available at: https://ssrn.com/abstract=3370917

Figure 6: The importance of Google data. True Real-time analysis with pre-selection of Google data.
RMSFE from: (i) models M1 - M13 with only variables extracted from Google data (in light gray), (ii)
models M1 - M13 with all the variables (St , IPt , laggedGDP and Google data) (in gray), (iii) models
N oGoogle1 - N oGoogle4 (in black).

Figure 7: Pseudo-Real-time versus True Real-time analysis (with pre-selection). Comparison of RMS-
FEs within the quarter from pseudo-real-time (in light gray) and true real-time (in gray) analysis. The
true real-time analysis does not include lagged GDP growth among the explanatory variables.

19

Electronic copy available at: https://ssrn.com/abstract=3370917

5 Conclusions
In this paper we consider the use of Google search data to nowcast the euro area
GDP growth rate. Our main objective is to evaluate the usefulness of Google search data
for nowcasting when official data are not available, against the background of a pseudo
real-time analysis. Because Google search data are high dimensional, in the sense that
the number of variable is large compared to the time series dimension, there is a price to
pay for using them: first, we need to reduce their dimensionality from ultra-high to high
by using a screening procedure, and second we need to use a regularized estimator to
deal with the pre-selected variables. Our second objective is to perform a true real-time
analysis and assess the validity of our results in this framework.

Four salient facts emerge from our empirical analysis. First, against the background
of a pseudo real-time analysis, we point out the usefulness of Google search data in
nowcasting euro area GDP growth rate for the first four weeks of the quarter when there
is no information about the state of the economy. We show that at the beginning of the
quarter, Google data provide an accurate picture of the GDP growth rate.

Second, as soon as official data become available, that is starting from week 5 with
the release of opinion surveys, then the relative nowcasting power of Google data instan-
taneously vanishes.

Third, we show that pre-selecting Google data before entering the nowcasting mod-
els appears to be a pertinent strategy in terms of nowcasting accuracy. Especially, we
implement the Sure Independent Screening approach put forward by Fan and Lv [2008]
enabling to retain only the Google variables that are the most correlated with the tar-
geted variable, that is GDP growth rate. This result confirms previous results obtained
with bridge equations augmented with dynamic factor (see e.g. Bai and Ng [2008] or
Schumacher [2010]).

Finally, we show when using Google search data in the context of a true real-time
analysis, the three previous salient facts remain valid. This result argues in favor of
the use of Google search data at the beginning of the quarter, when there is no official
information available, for real-time policy-making.

20

Electronic copy available at: https://ssrn.com/abstract=3370917

References
K. Aastveit and T. Trovik. Nowcasting norwegian GDP: the role of asset prices in a
small open economy. Empirical Economics, 42(1):95–119, 2012.

E. Angelini, G. Camba-Mendez, D. Giannone, L. Reichlin, and G. Ruenstler. Short-term

forecasts of euro area gdp growth. Economic Journal, 14:C25–C44, 2011.

J. Bai and S. Ng. Forecasting economic time series using targeted predictors. Journal of
Econometrics, 146(2):304 – 317, 2008. Honoring the research contributions of Charles
R. Nelson.

K. Barhoumi, O. Darne, and L. Ferrara. Are disaggregate data useful for forecasting
french gdp with dynamic factor models ? Journal of Forecasting, 29(1-2):132–144,
2010.

J. Boivin and S. Ng. Are more data always better for factor analysis? Journal of
Econometrics, 132:169–194, 2006.

M.-E. Bontempi, M. Frigeri, R. Golinelli, and M. Squadrini. Uncertainty, perception

and internet. Technical report, 2018.

D. Bragoli. Nowcasting the japanese economy. International Journal of Forecasting, 33

(2):390–402, April 2017.

D. Bragoli, L. Metelli, and M. Modugno. The importance of updating: Evidence from a

Brazilian nowcasting model. OECD Journal: Journal of Business Cycle Measurement
and Analysis, 2015(1):5–22, 2015.

D. Buono, G. Kapetanios, M. Marcellino, G. L. Mazzi, and F. Papailias. Big data

econometrics: Nowcasting and early estimates. Technical Report 82, Working Paper
Series, Universita Bocconi, 2018.

M. Carrasco and B. Rossi. In-sample inference and forecasting in misspecified factor

models. Journal of Business & Economic Statistics, 34(3):313–338, 2016.

Y. Carriere-Swallow and F. Labbe. Nowcasting with Google trends in an emerging

market. Journal of Forecasting, 32(4):289–298, 2013.

H. Choi and H. Varian. Predicting initial claims for unemployment insurance using
Google trends. Google Technical Report, 2009.

21

Electronic copy available at: https://ssrn.com/abstract=3370917

H. Choi and H. Varian. Predicting the present with google trends. Google Technical
Report, 2012.

D. Coble and P. Pincheira. Nowcasting building permits with Google Trends. MPRA
Paper 76514, University Library of Munich, Germany, 2017.

F. D’Amuri and J. Marcucci. The predictive power of Google searches in forecasting

unemployment. International Journal of Forecasting, 33:801–816, 2017.

M. Diron. Short-term forecasts of euro area real gdp growth: An assesment of real-time
performance based on vintage data. Journal of Forecasting, 27:371–390, 2008.

J. Fan and J. Lv. Sure independence screening for ultrahigh dimensional feature space.
Journal of the Royal Statistical Society B, 70:849–911, 2008.

L. Ferrara and C. Marsilli. Nowcasting global economic growth: A factor-augmented

mixed-frequency approach. The World Economy, 2018.

C. Frale, M. Marcellino, G. L. Mazzi, and T. Proietti. Euromind: a monthly indicator

of euro area economic conditions. Journal of the Royal Statistical Society, Series A,
174(2):439–470, 2010.

D. Giannone, L. Reichlin, and D. Small. Nowcasting: The real-time informational con-

tent of macroeconomic data. Journal of Monetary Economics, 55(4):665–676, May
2008.

D. Giannone, M. Lenza, and G. Primiceri. Economic predictions with big data: The
illusion of sparsity. mimeo, 2017.

T. Goetz and T. Knetsch. Google data in bridge equation models for german gdp.
International Journal of Forecasting, 35(1):45–66, 2019.

R. Golinelli and G. Parigi. Tracking world trade and GDP in real time. International
Journal of Forecasting, 30(4):847–862, 2014.

V. Kuzin, M. Marcellino, and C. Schumacher. MIDAS vs. mixed-frequency VAR: Now-

casting GDP in the euro area. International Journal of Forecasting, 27(2):529–542,
April 2011.

K.-C. Li. Asymptotic optimality of CL and generalized cross-validation in ridge regres-

sion with application to spline smoothing. The Annals of Statistics, 14(3):1101–1112,
1986.

22

Electronic copy available at: https://ssrn.com/abstract=3370917

K.-C. Li. Asymptotic optimality for Cp , CL , cross-validation and generalized cross-
validation: Discrete index set. The Annals of Statistics, 15(3):958–975, 1987.

X. Li. Nowcasting with big data : Is Google useful in the presence of other information?
mimeo, 2016.

M. Modugno, B. Soybilgen, and E. Yazgan. Nowcasting Turkish GDP and news decom-
position. International Journal of Forecasting, 32(4):1369–1384, 2016.

F. Narita and R. Yin. In search for information: Use of Google Trends’ data to narrow
information gaps for low-income developing countries. Technical Report WP/18/286,
IMF Working Paper, 2018.

P. Nymand-Andersen and E. Pantelidis. Google econometrics: Nowcasting euro area car

sales and big data quality requirements. Technical report, European Central Bank,
2018.

C. Schumacher. Factor forecasting using international targeted predictors: The case of

German GDP. Economics Letters, 107(2):95–98, May 2010.

S. Scott and H. Varian. Bayesian variable selection for nowcasting economic time series.
In A. Goldfarb, S. Greenstein, and C. Tucker, editors, Economic Analysis of the Digital
Economy, pages 119–135. NBER, 2015.

H. Varian. Big data: New tricks for econometrics. Journal of Economic Perspectives,
28(2):3–28, 2014.

S. Vosen and T. Schmidt. Forecasting private consumption: Survey-based indicators vs.

Google trends. Journal of Forecasting, 30(6):565–578, 2011.

23

Electronic copy available at: https://ssrn.com/abstract=3370917

Annex

Model Equation Predictors

0 (1) (1) (1)

M1 Yt = β0,1 + βg,1 xt,g + t xt,g = xt,g,1
0 (2) (2) (2) xt,g,1 +xt,g,2
M2 Yt = β0,2 + βg,2 xt,g + t xt,g = 2

0 (3) (3) (3) xt,g,1 +xt,g,2 +xt,g,3

M3 Yt = β0,3 + βg,3 xt,g + t xt,g = 3

0 (4) (4) (4) xt,g,1 +...+xt,g,4

M4 Yt = β0,4 + βg,4 xt,g + t xt,g = 4
(5) 0 (5) (5) (5) xt,1 +...+xt,5 (5)
M5 Yt = β0,5 + βs,5 xt,s + βg,5 xt,g + t xt,g = 5 , xt,s = St,1
(6) 0 (6) (6) (6) xt,1 +...+xt,6 (6)
M6 Yt = β0,6 + βs,6 xt,s + βg,6 xt,g + t xt,g = 6 , xt,s = St,1
(7) 0 (7) (7) (7) xt,1 +...+xt,7 (7)
M7 Yt = β0,7 + βs,7 xt,s + βg,7 xt,g + t xt,g = 7 , xt,s = St,1
(8) 0 (8) (8) (8) xt,1 +...+xt,8 (8)
M8 Yt = β0,8 + βs,8 xt,s + βg,8 xt,g + t xt,g = 8 , xt,s = St,1
(9) 0 (9) (9) (9) xt,1 +...+xt,9
M9 Yt = β0,9 + βs,9 xt,s + βg,9 xt,g + t xt,g = 9 ,
(9) St,1 +St,2
xt,s = 2
(10)0 (10) (10) (10) xt,1 +...+xt,10
M10 Yt = β0,10 + βs,10 xt,s + βg,10 xt,g + t xt,g = 10 ,
(10) St,1 +St,2
xt,s = 2
(11) (11)0 (11) (11) (11) xt,1 +...+xt,11
M11 Yt = β0,11 + βs,11 xt,s + βh,11 xt,h + βg,11 xt,g + t xt,g = 11 ,
(11) St,1 +St,2 (11)
xt,s = 2 , xt,h = IPt,1
(12) (12)0 (12) (12) (12) xt,1 +...+xt,12
M12 Yt = β0,12 + βs,12 xt,s + βh,12 xt,h + βg,12 xt,g + t xt,g = 12 ,
(12) St,1 +St,2 (12)
xt,s = 2 , xt,h = IPt,1
(13) (13)0 (13) (13) (13) xt,1 +...+xt,13
M13 Yt = β0,13 + βs,13 xt,s + βh,13 xt,h + βg,13 xt,g + t xt,g = 13 ,
(13) St,1 +...+St,3 (13)
xt,s = 3 , xt,h = IPt,1

Table 1: Equations of the 13 models (M 1, . . . , M 13) used to nowcast GDP growth over each quarter.
Equations include the variables pre-selected from Google data as well as information stemming from
surveys (St ) and industrial production (IPt ). St,i denotes the variable surveys St referring to the i-th
month of the current-quarter t and IPt,i denotes the growth rate of the industrial production available
at the 11th week of the current-quarter t and referring to the i-th month of the current-quarter t.

24

Electronic copy available at: https://ssrn.com/abstract=3370917

 Model Equation Predictors
(1) (1)
N oGoogle1 Yt = β0,1 + βs,1 xt,s + t xt,s = St,1
(2) (2)
N oGoogle2 Yt = β0,2 + βs,2 xt,s + t xt,s = St,1 +S 2
t,2

(3) (3) (3) (3)

N oGoogle3 Yt = β0,3 + βs,3 xt,s + βh,3 xt,h + t xt,s = St,1 +S
2
t,2
, xt,h = IPt,1
(4) (3) (4) (4)
N oGoogle4 Yt = β0,4 + βs,4 xt,s + βh,4 xt,h + t xt,s = St,1 +...+S
3
t,3
, xt,h = IPt,1

Table 2: Equations of the four models used to nowcast GDP growth without the variables extracted
from Google data. St,i denotes the variable surveys St referring to the i-th month of the current-
quarter t and IPt,i denotes the growth rate of the industrial production available at the 11th week of
the current-quarter t and referring to the i-th month of the current-quarter t.

25

Electronic copy available at: https://ssrn.com/abstract=3370917

 2 lags between estimation period and forecasting period
Last Training Nowcasting 1st GDP Vintage which contains Lagged week of new
Period Period the last GDP in the Training sample GDP Vintage
2013Q3 2014Q1 08/04/2014 no
2013Q4 2014Q2 08/04/2014 no
08/04/2014 no
15/04/2014 no 3rd week
04/06/2014 yes 10th week
2014Q1 2014Q3 02/07/2014 no
2014Q2 2014Q4 01/10/2014 no
21/10/2014 no 4th week
14/11/2014 yes 7th week
09/12/2014 yes 11th week
2014Q3 2015Q1 09/12/2014 no
17/03/2015 yes 12th week
2014Q4 2015Q2 17/03/2015 no
02/06/2015 yes 10th week
2015Q1 2015Q3 02/06/2015 no
30/07/2015 no 4th week
09/09/2015 yes 11th week
24/09/2015 yes 13th week
2015Q2 2015Q4 24/09/2015 no
13/11/2015 yes 7th week
08/12/2015 yes 11th week
2015Q3 2016Q1 08/12/2015 no
12/02/2016 yes 6th week
16/02/2016 yes 7th week
08/03/2016 yes 10th week

Table 3: Timeline of GDP release in real-time within the quarter. The first column gives the last
period used for the in-sample analysis (training sample), the second column indicates the nowcasting
period, the third column indicates the date of the first vintage which contains the GDP growth in the
last period of the training sample (indicated in the first column), the fourth columns indicates whether
a lagged GDP growth is available to be included among the explanatory variables (the corresponding
date and week of availability are given in the third and fifth columns, respectively). Finally, the fifth
column gives the week, and so the model, corresponding to the date in the third column.

26

Electronic copy available at: https://ssrn.com/abstract=3370917

 27

Electronic copy available at: https://ssrn.com/abstract=3370917

Figure 8: RMSFE from different models: “Google Data + Survey + IP” refers to models M1 - M13 with all the variables: St , IPt and Google
data, “Google Data” refers to models M1 - M13 with only variables extracted from Google data, “No Google Data” refers to models N oGoogle1
- N oGoogle4 , “Ridge + SIS Preselection” refers to model (2.2) estimated with pre-selected variables from Google data and Ridge regularization,
“Ridge” refers to model (2.2) estimated with Ridge regularization without pre-selection, “Google Data + Survey + IP + GDPlag” refers to models
M1 - M13 with all the variables: St , IPt , Google data and the lagged GDP growth when it is available, “True real time (w/o GDPlag)” refers to
models M1 - M13 estimated without including the lagged GDP growth in true real time.