Disponible à / Available at permalink :

(1)

- - -

Dépôt Institutionnel de l’Université libre de Bruxelles / Université libre de Bruxelles Institutional Repository

Thèse de doctorat/ PhD Thesis Citation APA:

Cicconi, C. (2012). Essays on macroeconometrics and short-term forecasting (Unpublished doctoral dissertation). Université libre de Bruxelles, Faculté Solvay Brussels School of Economics and Management, Bruxelles.

Disponible à / Available at permalink : https://dipot.ulb.ac.be/dspace/bitstream/2013/209660/6/08bf0a5a-fe36-4271-a626-aa467710eece.txt

(English version below)

Cette thèse de doctorat a été numérisée par l’Université libre de Bruxelles. L’auteur qui s’opposerait à sa mise en ligne dans DI-fusion est invité à prendre contact avec l’Université (di-fusion@ulb.ac.be).

Dans le cas où une version électronique native de la thèse existe, l’Université ne peut garantir que la présente version numérisée soit identique à la version électronique native, ni qu’elle soit la version officielle définitive de la thèse.

DI-fusion, le Dépôt Institutionnel de l’Université libre de Bruxelles, recueille la production scientifique de l’Université, mise à disposition en libre accès autant que possible. Les œuvres accessibles dans DI-fusion sont protégées par la législation belge relative aux droits d'auteur et aux droits voisins. Toute personne peut, sans avoir à demander l’autorisation de l’auteur ou de l’ayant-droit, à des fins d’usage privé ou à des fins d’illustration de l’enseignement ou de recherche scientifique, dans la mesure justifiée par le but non lucratif poursuivi, lire, télécharger ou reproduire sur papier ou sur tout autre support, les articles ou des fragments d’autres œuvres, disponibles dans DI-fusion, pour autant que :

Le nom des auteurs, le titre et la référence bibliographique complète soient cités;

L’identifiant unique attribué aux métadonnées dans DI-fusion (permalink) soit indiqué;

Le contenu ne soit pas modifié.

L’œuvre ne peut être stockée dans une autre base de données dans le but d’y donner accès ; l’identifiant unique (permalink) indiqué ci-dessus doit toujours être utilisé pour donner accès à l’œuvre. Toute autre utilisation non mentionnée ci-dessus nécessite l’autorisation de l’auteur de l’œuvre ou de l’ayant droit.

--- English Version ---

This Ph.D. thesis has been digitized by Université libre de Bruxelles. The author who would disagree on its online availability in DI-fusion is invited to contact the University (di-fusion@ulb.ac.be).

If a native electronic version of the thesis exists, the University can guarantee neither that the present digitized version is identical to the native electronic version, nor that it is the definitive official version of the thesis.

DI-fusion is the Institutional Repository of Université libre de Bruxelles; it collects the research output of the University, available on open access as much as possible. The works included in DI-fusion are protected by the Belgian legislation relating to authors’ rights and neighbouring rights.

Any user may, without prior permission from the authors or copyright owners, for private usage or for educational or scientific research purposes, to the extent justified by the non-profit activity, read, download or reproduce on paper or on any other media, the articles or fragments of other works, available in DI-fusion, provided:

The authors, title and full bibliographic details are credited in any copy;

The unique identifier (permalink) for the original metadata page in DI-fusion is indicated;

The content is not changed in any way.

It is not permitted to store the work in another database in order to provide access to it; the unique identifier (permalink) indicated above must always be used to provide access to the work. Any other use not mentioned above requires the authors’ or copyright owners’ permission.

(2)

U niversité L ibre de B ruxelles

Solvay Brussels School of Economies and Management

“Essays on macroeconometrics and short-term forecasting”

Dissertation présentée en vue de l'obtention

du grade de Docteur en Sciences économiques et de gestion par

Claudia Cicconi

sous la direction du Professeur Domenico Gîannone

B oo

(3)

Acknowledgments

Completing this thesis has been without any doubt the most absorbing expé

rience of my life: I think that no single day has passed during the last years in which I hâve neither worked on it nor thought that I should hâve. De- ciding to conclude the PhD while maintaining my full time job in Italy has certainly made things much more difficult. I hâve several times asked my self WH Y?’ or SURE??’ but after weeks of complété darkness I hâve every time found, somehow, the motivation to go on.

I owe this resuit to many people that helped me along the way. First of ail, I want to thank, even if ’to thank’ is definitely not enough, my family who supported me in an amazing and endearing way. I then want to thank, from the heart, Luca Sala for his help at the very beginning of this expérience;

Lucrezia for her words when I told her that I had decided to leave Brussels and the PhD and for her trust and support afterwards; Maxjoree and Georg for encouraging me these last years. I also thank Saverio Simonelli, with whom I worked on the first chapter of the thesis, for fruitful, stimulating and enriching collaboration and the members of the Jury whose comments and suggestions hâve improved the thesis. And last but certainly not least, I want to thank Domenico for his invaluable research guidance and for being a fantastic coaeh, managing to give me the right motivation to overcome the most insidious obstacles. I am also very grateful to my colleagues in ISTAT and in particular to Marco, Francesca and Ludovico for their support and compréhension these last years. And also to Raffaella, Pablo, Cecilia and Augusto for their maxvelous practical and psychological help these last days.

It has been very hard but I hâve been very lucky. I spent in Brussels wonderful years. Many people helped me to make life easier and better: from the EGARES staff, Claude in particular, to Alexis - my colleghetto - Fulvio, Jacopo, Sergey, Antonello, Ugo, Gianluca, Luca, David - my Nutella friend - Alexandre, Cristina, Lollo, Joachim, Maria Caterina, Cristina, Ghislaine ...

Some other made the big différence: Paolo, Alexander, Chiara and Micol,

(4)

Laura, Prancesca, Paulo, Pablo, Giovanna and Marcello .... and I am happy that most of them still make it despite geography and air companies play against.

I feel very lucky also for meeting very spécial people in Rome. I want to thank old and new friends for sharing with me not only wonderful carefree mo

ments but also, and foreniost, tlie difficult ones. Auguste, Cecilia, Margherita, Ciccio, Giordana, Annachiara, Daniela, Barbara, Andrea, Enrico, Rafîaella, Federica, Sabina, Alessandra, Domenico and ail the others....! hope you will keep on being an important part of my life for a long time in the future.

La fine è l’inizio!

(5)

List of Tables

1.1 The release calendar of Italian data in 2009 ... 44

1.2 Data availability patterns... 45

1.3 Forecast uncertainty over the whole évaluation period... 47

1.4 Forecast uncertainty over the crisis... 47

1.5 Forecast uncertainty with anticipated IP releases... 50

1.6 Cut-üff dates of professional and institutional forecasters .... 51

2.1 Data set... 68

2.2 Observed release calendar and timing of the updates... 69

2.3 Unbalance patterns (months)... 70

2.4 Relative MSFE of the employment growth rate of the four largest Member States... 73

3.1 Not modeling the block structure = 1, n*’ = 10)... 99

3.2 Not modeling the block structure (r*^ = 3, n** = 10)... 100

3.3 Not modeling the block structure (r*^ = 1, n*’ = 20)... 101

3.4 Miss-specifying the block structure. Effects ou conunon factors 102 3.5 Speed of convergence of the estimation procedure...103

(8)

List of Figures

1.1 GDP, the industrial production index and the Economie Senti

ment indicator ... 46 1.2 Dynamic corrélation... 46 1.3 Box plots of noweasts’ errors over the whole évaluation period . 48 1.4 Box plots of noweasts’ errors over the crisis... 49 1.5 MSFEs of q-o-q growth rates of GDP... 52 1.6 Noweasts of current year GDP annual growth rate... 53 1.7 MSFEs of current year GDP annual growth rates noweasts . . 54 2.1 Mean square forecast errors of the estimâtes of quarter q made

in g — 1, g and g + 1... 71 2.2 Estimâtes of the current quarter ... 72

(9)

Introduction

The thesis, entitled "Essays on macroeconometrics and short-term forecasting”, is composed of three chapters. The first two chapters are on nowcasting, a topic that has received an increasing attention both among practitioners and the academies especially in conjunction and in the aftermath of the 2008-2009 économie crisis. At the heart of the two chapter is the idea of exploiting the information from data published at a higher frequency for obtaining early esti

mâtes of the macroeconomic variable of interest. The models used to compute the noweasts are dynamic models conceived for handling in an efficient way the main characteristics of the data used in a real-time context, like the fact that due to the different frequencies and the non-synchronicity of the releases the time sériés hâve in general missing data at the end of the sample. While the first chapter uses a small model like a VAR for nowcasting Italian GDP, the second one makes use of a dynamic factor model, more suitable to handle medium-large data sets, for providing early estimâtes of the employment in the euro area. The third chapter develops a topic only marginally touched by the second chapter, i.e. the estimation of dynamic factor models on data characterized by block-structures.

The first chapter assesses the accuracy of the Italian GDP noweasts based on a small information set consisting of GDP itself, the industrial production index and the Economie Sentiment Indicator. The taak is carried out by using real-time vintages of data in an out-of-sample exercise over rolling Windows of data. Beside using real-time data, the real-time setting of the exercise is also guaranteed by updating the noweasts according to the historical release calendar. The model used to compute the noweasts is a mixed-frequency Vec- tor Autoregressive (VAR) model, cast in state-space form and estimated by maximum likelihood. The results show that the model can provide quite accurate early estimâtes of the Italian GDP growth rates not only with respect to a naïve benchmark but also with respect to a bridge model based on the same information set and a mixed-frequency VAR with only GDP and the

(10)

industrial production index.

The chapter also analyzes with some attention the rôle of the Economie Sen

timent Indicator, and of soft information in general. The comparison of our mixed-frequency VAR with one with only GDP and the industrial production index clearly shows that using soft information helps obtaining more accurate early estimâtes. Evidence is also found that the advantage from using soft information goes beyond its timeliness.

In the second chapter we focus on nowcasting the quarterly national ac- coimt employment of the euro axea making use of both country-specific and area wide information. The relevance of anticipating Eurostat estimâtes of employment rests on the fact that, despite it represents an important macroeconomic variable, euro axea employment is measured at a relatively low frequency (quarterly) and published with a considérable delay (approximately two months and a half ). Obtaining an eaxly estimate of this variable is possible thanks to the fact that several Member States publish employment data and employment-related statistics in advance with respect to the Eurostat release of the euro area employment. Data availability represents, nevertheless, a major limit as country-level time sériés are in general non homogeneous, hâve different starting periods and, in some cases, are very short. We construct a data set of monthly and quarterly time sériés consisting of both aggregate and country-level data on Quarterly National Account employment, employment expectations from business surveys and Labour Force Survey employment and unemployment. In order to perform a real time out-of-sample exercise simulating the (pseudo) real-time availability of the data, we construct an artificial calendar of data releases based on the effective calendar observed during the first quarter of 2012. The model used to compute the nowcasts is a dynamic factor model allowing for mixed-frequency data, missing data at the beginning of the sample and ragged edges typical of non synchronous data releases. Our results show that using country-specific information as soon as it is available allows to obtain reasonably accurate estimâtes of the employment of the euro area about fifteen days before the end of the quarter.

We also look at the nowcasts of employment of the four largest Member States. We find that (with the exception of Fiance) augmenting the dynamic factor model with country-specific factors provides better results than those obtained with the model without country-specific factors.

The third chapter of the thesis deals with dynamic factor models on data characterized by local cross-correlation due to the presence of block-structures.

The latter is modeled by introducing block-specific factors, i.e. factors that are spécifie to blocks of time sériés. We propose an algorithm to estimate

(11)

the model by (quasi) maximum likelihood and use it to run Monte Carlo simulations to evaluate tlie effects of modeling or not tlie block-structure on the estimâtes of common factors. We find two main results: first, that in finite samples modeling the block-structure, beside being interesting per se, can help reducing the model miss-specification and getting more accurate estimâtes of the common factors; second, that imposing a wrong block-structure or imposing a block-structure when it is not présent does not hâve négative effects on the estimâtes of the common factors. These two results allow us to conclude that it is always recommendable to model the block-structure especially if the characteristics of the data suggest that there is one.

(12)

Chapter 1 Nowcasting Italian GDP in real-time

Abstract: This chapter provides nowcasts of the Italian GDP growth rate using a small information set consisting of GDP itself, the industrial produc

tion index and the Economie Sentiment Indicator. The task is carried out by performing a real-time out-of-sample exercise that allows to analyze the évolution of forecast uncertainty as additional information cornes available ob- serving the real-time historical availability of the data. The data set consists of real-time vintages of time sériés and the model used to compute the now

casts is a mixed-frequency Vector Autoregressive model, put in state-space form and estimated by maximum likelihood using the Expectation maximiza- tion algorithm.

Keywords: mixed-frequency Vector Autoregressive model, state-space form, Kalman filtering and smoothing, mixed-frequency data, ragged-edge data.

Economie Sentiment Indicator, soft information JEL: E52, C33, C53

This chapter is part of an ongoirig work with Saverio Siinonelli (European University In- stitutc and University of Naples Federico II). Wc wish to thank, without irnplicating them, participants to the 6th Eurostat Colloquium on Modem Tools for Business Cycle Analysis (Luxembourg, Septernber 2010) and to the 30th CIRET Conférence on Tendency Surveys and the Service Sector (New York, October 2010).

(13)

1.1 Introduction

Policy makers, central bankers, private banks and financial operators are called to take strategie decisions on the basis of their knowledge of the current State of the economy. This task is made barder by the publication lag that charac- terizes some key économie indicators, especially those recorded at quarterly frequency. Producing flows of frequently and regularly updated noweasts of quarterly économie indicators before they are officiaJly published is therefore extremely relevant (see, among others, Giannone, Reichlin, and Small, 2008 and Bahbura, Giannone, and Reichlin, 2010).

The key ingrédient of noweasting is the use of statistical information that is related to the économie Vctriable to noweast but released with a shorter delay and/or at a higher frequency - monthly, weekly or even daily. Translating this information into a reliable estimate of an official quarterly macroeconomic variable, however, is not an easy task. What is crucial, in addition to the choice of the data, is the use of a suitable econometric model. The latter, in fact, should be able to incorporate efficiently the most up to date information and to handle data of varions periodicity, releeised in a non-synchronous manner and with non-homogeneous publication lags.

One of the most used simple noweasting model, largely used by central banks, is the bridge équation model (see, e.g. Trehan, 1989 and Parigi and Schlitzer, 1995). The main idea of the bridge équation model to link, or 'bridge’, the informational content of monthly indicators to the quarterly target variable is, in fact, at the basis of noweasting. However, the bridge équation may resuit too simple to capture the real-time available informa

tion. Recent advances in the literature brought to the formulation of more complex econometric models, more effective at handling the characteristics of the data used for noweasting: mixed-frequencies, potential missing data at the beginning of the sample for some time sériés and ragged edges typical of non-synchronous data releases. These econometric models are essentially of three types: mixed frequency VAR (Vector autoregressive) models and MI- DAS (Mixed DAta Sampling) models - for data sets of small size - and mixed frequency dynamic factor models for medium-to-large data sets.

In this chapter we présent the noweasts of the quarter-on-quarter growth rate of Italian GDP based on a mixed-frequency VAR in which the dynamics of the quarterly target variable is jointly modeled with two monthly indicators:

the industrial production index (IPI) and the Economie Sentiment Indicator (ESI). The methodological framework under which the model is developed is the same as in Bahbura, Giannone, and Reichlin (2010). The model is written

(14)

in state-space form and the Kalman filter is used to deal with missing data arising from mixed frequencies and ragged edges at the end of the sample and to estimate the model by maximum likelihood via the EM algorithm.

The choice of a VAR model restricts our attention to small size models for nowcasting GDP. Such a choice is justified by three main considérations.

Firstly, the data set we use is small but extremely informative on GDP as the Italian estimate of GDP is basically a supply-side estimate and, being industry the sector characterized by the most pronounced cyclical dynamics, it is mainly determined by the IPI. Secondly, limiting the attention to small size models, and in particular to mixed frequency VAR, allows for a direct comparison with the bridge équation model which is one of the nowcasting models most used in practice, especially in Italy. Indeed, the mixed-frequency VAR can be seen as an évolution of the bridge équation model: it shares with the latter the same philosophy of using monthly variables to get information on the quarterly target variable but it allows to jointly model the dynamics of monthly and quarterly variables and, being a multivariate model, allows for feedbacks from quarterly to monthly variables that are excluded in the bridge univariate model. Thirdly, limiting the attention to three variables allows us to use a genuine real-time data set to evaluate the nowcasts.

The main contribution of the chapter consists of evidence on the accuracy of Italian GDP nowcasts based on a real-time out-of-sample exercise.

The latter is conducted by letting the real-time vintages of GDP, the IPI and the ESI enter the model according to their historical release calendar, therefore reproducing a perfect real-time situation. This is, at least to our knowledge, the only experiment in this sense with Italian data. Golinelli and Parigi (2008), in fact, used a real-time data set only for GDP, while the time sériés of the monthly indicators were revised data to which seasonal adjustment was applied in a real-time fashion. Moreover, while Golinelli and Parigi (2008) assume their nowcasts are made at the end of each month, we update the nowcasts more often in order to analyze the informational content of the release of each variable. The real-time out-of-sample exercise shows that the uncertainty of the nowcasts obtained with our model significantly reduces as additional information coming from new data releases is incorporated in the model, confirming the ability of the mixed-frequency VAR to incorporating relevant up to date information. The model is found to outperform not only a benchmark constant growth model but also two other competitors: a mixed- frequency VAR with only GDP and the IPI and a bridge équation model with the same input variables as the tri-variate VAR. Particular attention is also devoted to assessing the rôle of soft data at improving nowcast accuracy. The évaluation is carried out not only by comparing our VAR with the VAR with

(15)

only GDP and industrial production but also by simulating two alternative release scénarios in which the publication lag of the IPI is reduced. The sec

ond experiment allows us to draw some conclusions on whether the ESI has some informational content beyond its timeliness.

The chapter is structured as follows. Section 1.2 provides a review of alternative approaches to nowcasting proposed in the literature; section 1.3 présents the mixed-frequency VAR used in the out-of-sample exercise; section 1.4 describes the main characteristics of the data and of their release calendars;

section 1.5 focuses on the design of the out-of-sample exercise, given the release calendar of the data. Section 1.6 présents the results and section 1.7 concludes.

1.2 Related literature

The publication lag of the main économie indicators has progressively reduced along time. The publication lag of Italian Quarterly National Accounts (QNA) has reduced from between 90 and 120 days in 1984 to 70 days in 2003 and from November 2000 a flash estimate of GDP is released with a delay of 45 days. Nowcasting, therefore, was more relevant in the past than it is in the présent. Yet, for a long time it has been practiced only informally or using very simple econometric models and received the attention of the academie literature only recently.

One of the most used informai nowcasting procedure implemented at in

ternational organizations and policy institutions is the so called judgemental forecasting: experts are called to formulate, according to a ’model’ they only hâve in their mind, an estimate of one or more macroeconomic variables of interest based on the most up to date information available. Judgemental forecasting is extremely fascinating as it rests almost exclusively on the skill and the expérience of the expert and takes into account both statistical and non statistical information, but suffers from several limits: firstly, it is ex

tremely subjective; secondly, the ’model’ in the expert mind is unknown and potentially unstable over time; thirdly, it is very diflflcult if not impossible for the expert to report with some rigor why e.g. an estimate is revised 0.2 up or down (and why 0.2 and not 0.3 or 0.1). For these reasons judgmental forecasting is most often used only to complément model based forecasts.

The most used simple econometric model for nowcasting, still implemented at varions central banks to noweast GDP and other QNA aggregates, is the so called bridge équation model (see e.g. Trehan, 1989). A bridge équation model is nothing else than a régression of the target quarterly variable on the

(16)

quarterly aggregates of one or more monthly indicators. The régression allows to 'bridge’ the monthly information that cornes available along the quarter to information on the quarterly variable of interest. We can distinguish two types of bridge models: in ’supply-side’ models, GDP is directly nowcast via a single-equation approach (see Trehan, 1989) using supply-side indicators; in

’demand-side’ models, instead, GDP is derived indirectly as the sum of its demand-side components, each obtained by using one spécifie bridge équation (see e.g. Parigi and Schlitzer (1995)). Other applications of bridge models can be found in Kitchen and Monaco (2003), Rünstler and Sédillot (2003), BafRgi, Golinelli, and Parigi (2004) and Diron (2008) for the euro area GDP, in Parigi and Golinelli (2007) for the G7 countries and the EU and in Golinelli and Parigi (2008) for Italy. The empirical evidence emerging from these contri

butions allows to conclude that using bridge équations improves the accuracy of GDP noweasts. However, the advantage over benchmark models without monthly indicators is limited to situations in which the monthly indicators are known for the whole quarter to noweasts or only the last month is missing. This may be due to one of the major limits of bridge models, i.e. the need of relying on auxiliary models to forecast the missing observations of the monthly indicators at the end of the sampleU A further but related limit of this approach is that, like any other single équation approach, no feedback is allowed among monthly indicators and from the quarterly variable to the monthly indicators.

The literature has recently proposed more sophisticated econometric mod

els than the bridge équations, suited to deal with the characteristics of the data used for noweasting: i) data may hâve different frequencies, ii) time sériés may start at different points in time (missing observations at the beginning of the sample), iii) time sériés may end at different points in time due to their different periodicity and non-synchronous data releases (ragged edges), iv) the number of time sériés may be potentially large.

Dynamic factor models for mixed frequencies hâve been proposed for now

easting by Evans (2005) and Giannone, Reichlin, and Small (2008), whose approach was first implemented at the Board of Governors of the Fédéral Re

serve. Similar models are also implemented at the European Central Bank^

and at several national central banks^. Proietti (2008) proposed a large dy- *

*The auxiliary models employed to forecast the monthly indicators vary across papers from autoregressive models to VARs and Bayesian VARs. Most often they arc not clearly reported by the authors.

^See for e.xample Angelini et al. (2008) and Baribura and Rünstler (2007).

^See, among others, D’Agostino et al. (2008) for Ireland and Matheson (2010) for New Zcaland.

(17)

namic factor model to forecast euro area GDP and its components (17 quarterly variables) by using more than hundred monthly indicators. Proietti (2008)’s model, however, cannot be used in a real-time setting as it cannot liandle ragged edge time sériés. Banbura et al. (2010) further developed the approach in Giannone, Reichlin, and SmaJl (2008) introducing maximum likelihood estimation of mixed frequency factor model and proposing a model- based approach to measure the impact of news in data releases on the nowcasts. Banbura and Modugno (2010) developed the approach outlined in Giannone, Reichlin, and Small (2008) and Banbura et al. (2010) proposing a more general methodological framework to model the joint dynamics of (possibly many) time sériés characterized by mixed frequencies and ragged edges, also considering missing data at the beginning of the sample. Their approach is general in the sense that can be applied to any model that can be written in state-space form'*.

While dynamic factor models can be used to handle medium-to-large data sets, Ghysels, Santa-Clara, and Valkanov (2004) introduced Mixed Frequency DAta Sampling (MIDAS) models, a univariate régression model suited for nowcasting with a few indicators^. In the MIDAS régression the (target) lower frequency variable is regressed on distributed lags of one or more indicators sampled at a higher frequency. The innovation with respect to traditional distributed lag models (see e.g. Koenig et al., 2003) consists of expressing the coefficients of the lag pol)momials of the high frequency indicators as a fonc

tion of a very small-dimensional vector of parameters®. MIDAS régressions, originally conceived for other purposes, were first applied to nowcast macroeconomic data by Cléments and Galvao (2006), Ghysels and Wright (2006) and Ghysels and Valkanov (2006). Marcellino and Schumacher (2008) and Fraie and Monteforte (2010) combined dynamic factor models and MIDAS régressions: the former by using estimated factors as regressors in the MI

DAS équation, the latter by including MIDAS polynomials of the indicators as items in a dynamic factor model. Bai, Ghysels, and Wright (2010) parallel state-space approaches in general to MIDAS.

An alternative approach to nowcasting is provided by mixed frequency

"*366 Banbura and Modugno (2010) and the references therein for details on the methodology and the estimation procedure.

®The idea of MIDAS has bcen developed and extended in scveral directions in the last few years. For a survey see e.g. Andreou, Ghysels, and Kourtcllos (2010) and the references therein.

®Two flexible spécifications that parameterize the weights into a two parameter vector are the Beta lag and the two-parameters exponential Almon lag. Sec e.g. Ghysels, Santa-Clara, and Valkanov (2004).

(18)

Vector autoregressive models. Mixed-frequency VAR hâve been introduced by Zadrozny (1990) and further developed by Mittnik and Zadrozny (2004) and Hyung and Granger (2008). Mixed-frequency VARs, like traditional VARs, can only be used to model a small number of time sériés. In order to cope with the high dimensionality of parameters when larger data sets are considered, Bayesian mixed frequency VARs hâve been recently proposed in the literature.

One of the most prominent contribution in this field is by Schortfeide and Song (2011), who developed a mixed frequency VAR estimated under Minnesota- priors to forecast GDP using two quarterly and eight monthly variables^. An alternative Gibbs sampling approach with respect to Schortfeide and Song (2011)’s is explored in Ghiu et al. (2011), though the interest of their paper is more on estimation than on forecasting. A mixed frequency VAR estimated by maximum likelihood has been used by, e.g., Giannone et al. (2009) to nowcast euro area GDP.

Mixed-frequency VARs are very much related to bridge équation models, with respect to which they can be seen as a generalization. In fact, a mixed- frequency VAR; i) allows to model the joint dynamics of time sériés sampled at different frequencies; ii) can deal with time sériés with ragged edges, without requiring the use of auxiliary models to forecast the missing observations of the indicators at the end of the sample; iii) is a multivariate model.

Mixed frequency VARs are also related to MIDAS models. They can both tackle the mixed-frequency nature of the data and exploit the observations of timely indicators available at higher frequency, but they also show marked dif

férences: i) MIDAS is a single équation approach whereas the mixed-frequency VAR is a multivariate model; ii) the mixed-frequency VAR sufîers from the curse of dimensionality while the MIDAS model allows for a parsimonious parametrization of the coefficients. However, the MIDAS restrictions on the lag polynomials may be invalid while the VAR has the advantage of being es

timated without parameter restrictions. Ghysels and Valkanov (2006) showed that a MIDAS régression can be seen as an approximation of a general dynamic linear model, possibly a VAR, where the low frequency variable is a stock variable. In principle, therefore, if the data generating process is a VAR the mixed-frequency VAR should perform better than the MIDAS, but which of the two models performs better in practice is essentially an empirical issue®.

Our nowcasts are derived from a mixed-frequency VAR with three vari-

^See Schortfeide and Song (2011) for greater details on the model, the estimation method and the empirical results.

®See e.g. Kuzin, Marcollino, and Schumacher (2009) for a cornparison between MIDAS régressions and mixed-frequency VARs.

(19)

ables: GDP and two monthly indicators - the IPI and the ESI. The model is developed in the same general nowcasting framework introduced by Giannone, Reichlin, and Small (2008) for large mixed frequency factor models and it is estimated using the general techniques proposed by Bahbura and Modugno (2010). The model, in fact, is cast in state-space form and the Kalman filter is used to handle missing data and to estimate the model by maximum likelihood with the EM algorithm.

1.3 The model

The VAR is in general a suitable tool for modeling small information sets as it is able to capture the linear dynamic interactions among the variables of interest. The mixed-frequency VAR used in this chapter is a generalization of the traditional VAR that allows to take account of the characteristics of the data used for nowcasting in real-time:

- mixed-frequencies: we want to model a quarterly variable jointly with selected indicators sampled at monthly frequency;

- ragged edge data; as the data we consider are characterized by different publication lags and non synchronous releases the time sériés at any given point in time are not balanced, showing a different number of missing observations at the end of the sample

A mixed frequency VAR is a VAR operating at the highest sampling fre

quency of the time sériés included in the model (see (Zadrozny, 1990), (Mit- tnik and Zadrozny, 2004) and (Hyung and Granger, 2008)). Low frequency variables are treated as partially unobserved high frequency variables that re

spect proper time aggregation restrictions in accordance with their stock/flow nature. The high frequency VAR and the time aggregation restrictions can be cast in state-space form and estimated by maximum likelihood. In this framework, Kalman filtering techniques are used to handle ragged edges and to take account of the mixed frequency nature of the data.

Let be the quarterly variable - GDP in our case - and Xt^ the k x 1 vector of monthly indicators (fc = 2 in our case)®. For setting the time aggregation constraint we follow Maxiano and Murasawa (2003). In particular, it is assumed that the quarterly variable has a unit root^® and satisfies the

®The number of quarterly variable is assumed to be one but it can be larger, without any loss of generality.

*°The augmented Dickey Fuller test against the alternative of trend stationarity for GDP cannot reject the null hypothesis of a unit root at the 10 per cent level.

(20)

relation

iny,^ = iny;^ + iny,:^_i + iny;^_2, = 3,6,...,t„. (i.i) Then, taking the three period différences of in (1.1)

iny„-iny„_3 = (iny;„-iny;_3) + (iny;^_i-iny,;_4) +(iny,;_2-iny;_5)

and hence, defining yt^ = A3 In Yt„ the observed q-o-q growth rate of the quarterly variable, and = Alny^’]^, the m-o-m growth rate of the corresponding unobserved monthly sériés,

Htm = {ytm + Vtm-l + + 2/*,„-2 + Vlm-s) + yt„-3 + yln-4)

= 2/*„ + 2y*^_i + iyl^-2 + 22/r„_3 + Vu-4 (1-2) which holds for tm = 3,6,9,..., Tm because GDP is assumed to be only ob

served every third month of each quarter.

Then, the unobserved month-on-month growth y*^ and the monthly indicators Xt^ are assumed to follow the VAR{p)

A{L) yL - y*y - yx with A{L) = and et^ ~ i.i.d.N{0,'Z).

(1.3)

The VAR{p) (1.3) and the time aggregation constraint (1.2) can be cast in state-space form. Let us assume p ^ 4^^ and define the State vector

2«m-4

with zt^ = ytm - yy

^tm - yx (1.4)

Writing the VAR{p) (1.3) in companion form, combining it with the time aggregation constraint (1.2) and setting

yL

-{

ytm ^rn — 3,6,..., Tm 0( otherwise

"The statc-spacc représentation for p > 4 can be dérivée! in a straightforward manner by modifying the State vcctor and the System matrices accordingly.

(21)

where at^ is a random draw from a distribution that does not dépend on the parameters of the model, the State space-form of the mixed-frequency VAR{p) is obtained as

where

yt - y-t

- yx = BtmStm + (1.5)

«t„+i = + Dvt ⁽¹^.⁶⁾

t-m — 3,6,..., Tm p

0 otherwise 0 otherwise

0 tm = 3,6,..., Tm v-i/2

^ otherwise and the System matrices are:

B O'z , 2kxl 0^1

Or

C

kxl -‘A:

Cl

C2 , Cl = [ (fl

3 0^1 2 0'- 1

O'^â Ofexl Ofcxl Ofexl O^âxi Ofcxl 0-k

Tp ^kxk{5--p) ] > <^2 = [ hic 0.ikxkk J >

where k = k + 1.

D E1/2 ■

04fexfc .

The mixed-frequency VAR{p) is estimated by maximum likelihood with the EM algorithm. See Bahbura and Modugno (2010) and Mariano and Mura- sawa (2003) for details on the dérivation of the expectation and the maximiza- tion steps of the algorithm. Once the parameters are estimated, the forecasts of the q-o-q growth rate of the quarterly variable are obtained by applying the Kalman smoother, ensuring that ail the timely observations from the monthly indicators are taJcen into account.

Mariano and Murasawa (2003)’s approach ensures the line^ity of the state-space représentation of the model, thereby allowing to apply standard Kalman filtering techniques. The disadvantage of this approach is that it does not ensures the consistency between the observed levels of the quarterly vari

able and the quarterly levels obtained by time aggregating the corresponding

(22)

unobserved monthly sériés. This is because the time aggregation constraint is derived by assuming

+ inr,:_i + inr;„_2- (1.7) instead of the exact time aggregation constraint

= + (1.8)

The latter, at the basis of the so called ’cumulator variable’ approach (see e.g. Harvey, 1991 and Harvey and Chung, 2000), ensures consistency between monthly and quarterly levels but implies a non-linear state-space représen

tation that reqtiires to iterate the Kalman filter and smoother adapted to a sequentially linearized state-space form (see e.g. (Evans, 2005) and (Proietti, 2008)), which is rather troublesome from a computational point of view.

Here we use Mariano and Murasawa (2003) to préservé the linearity of the Kalman filter and because, despite the lack of consistency in levels, the approximation provided by the time aggregation constraint works quite well for growth rates, i.e. what we are interested in. The same approach is followed, among others, by Kuzin et al. (2009), Giannone et al. (2009), Bahbura and Modugno (2010), Bahbura et al. (2010).

1.4 The real-time data set

Our data set consists of real-time vintages of time sériés of GDP, the industrial production index and the Economie Sentiment Indicator. The former two are published by the Italian National Institute of Statistics (ISTAT), while the latter is published by the Direction General of Economie and Financial Affairs (DG ECFIN) of the European Commission.

The flash estimate of Italian GDP is based on the same accounting scheme as the whole System of QNAs^^. Such a scheme is a very complex one, trying to replicate as close as possible the methodology used for annual National Accounts, and it is based on a very large number of économie indicators considered at a very disaggregate level. Replicating the informational content of the QNA estimâtes with such a small model as a VAR is obviously not even thinkable. However, we daim, and our results confirm, that the combination

'^In some countries flash estimâtes reflect simplifled accounting schemes or ad hoc econo- inetric inodcls that inake use of a very reduced set of information with respect to the revised estimâtes.

(23)

of the data and the model we use conveys an absolutely relevant amount of information for nowcasting Italian official G DP data.

The accuracy of our nowcasts is strictly related to the fact that the IPI is a key déterminant of GDP. The main reason why it is so is that the estimate of Italian GDP is basically a supply-side estimate and, since industry has a marked cyclical dynamics compared to the other sectors of the economy, the IPI is the main piece of information on which the GDP estimate is based.

GDP, in fact, is released only after the IPI referred to the last month of the quarter to estimate is available.

In addition to the IPI we include in the data set the Economie Sentiment Indicator, that revealed to be, on the basis of a preliminary analysis of ail the monthly business and consumer surveys (B&C) indicators, the most robust and effective survey indicator to complément the information gathered by the IPP^. It is well known that, due to their qualitative nature, business and consumer surveys are usually reckoned as less informative than hard data.

Nonetheless, they are often used for nowcasting because of their extremely reduced publication lag (see e.g. Giannone, Reichlin, and Small, 2008).

Once given some qualitative information on the composition of the data set used, we provide some more technical details on its real-time characteristics.

1.4.1 Gross Domestic Product

The GDP data set consists of the real-time vintages of time sériés of the seasonally (and calendar) adjusted chained linked volumes of GDP as issued by the Italian National Institute of Statistics (ISTAT) from February 2001 (February 15, 2001) to December 2010 (December 10, 2010). Italian quarterly GDP data are released eight times a year: in February, March, May, June, August, September, November and December. The releases in February, May, August and November concern flash estimâtes of GDP of the previous quarter, while those in March, June, September and December concern first revised estimates^^.

'^In a preliminary study we included in the rnixed-frequency VAR, one in turn, the bal

ances of opinions of ail the monthly questionnaires related to ail of the sectors of activity (except services) and families, the consumer confidence indicator and the confidence indica

tors of ail the sectors (except services). None of the indicators performed particularly well compared to the Economie Sentiment Indicator.

'■^Three vintages arc missing in the data set, corresponding to three occasions in which data hâve not been published: February 2006 (flash estimate of the last quanter of 2005), February and March 2008 (flash and first revised estimâtes of the last quarter of 2007, respectively).

(24)

Since 2001 Italian QNA hâve undergone two major révisions: the first one in June 2003 and the second one in March 2006 (benchmaxk revision^^). In ad

dition to extraordinary révisions, GDP (and the whole QNAs) is also revised every year in March to incorporate the new estimâtes of the annual national accounts released a few days earlier. National account annual data are kept fixed between two March estimâtes of QNA: e.g., QNA estimâtes published on March 10 2010 incorporated the annual estimâtes of the National Accounts from 2007 to 2009 that were released ten days before. These estimâtes were then used to produce GDP and QNA estimâtes from March 2010 to February 2011 (flash GDP estimate of GDP of 2010:q4); then, in March 2011 the QNAs incorporated the estimâtes of the annual National Accounts for 2010, 2009 and 2008 published on March 1 2011, that hâve been kept fixed until Febru

ary 2012. The révisions of QNA that occur in March are in general larger than the ordinary révisions that characterize ail the releases. The latter are mainly due to the updating of the indicators used in the estimation procedure and to the statistical treatment of the data (namely, temporal disaggregation and calendar and seasonal adjustment). It is worth noting that, though these révisions hâve an efîect on the whole time sériés, only the most recent obser

vations (normally the two previous complété years plus the quarters of the current year) are revised between two March releases.

1.4.2 The industrial production index

The data on industrial production consists of 118 vintages of seasonally (and calendar) adjusted time sériés from January 1995'®, published from March 2001 (March 15, 2001) to December 2010 (December 10, 2010).

The vintages are characterized by two exceptional révisions of raw data.

The first one occurred in March 2003 for the adoption of the new classifica

tion of économie activities (Nace Rev.1.1) and the change of the base year from 1995=100 to 2000=100; the second one occurred in March 2009, for the adoption of the Nace Rev.2 and the change of the base year from 2000=100

'^Thc tcrm benchmark revision is used in relation to National Aecounts to indicate an ex- eeptional révision due to the adoption of major definitional and metliodological innovations.

The last benchmark révision of Italian QNA occurred in December 2011 to adéquate the QNA to the new classification of the économie activities (.Nace Rev.2) already introduced in other domains (like e.g. that of short-term indicators).

'®1STAT officially publishes the rcal-time vintages of the IPI time sériés starling in .January 2000 and released from March 2001. The period January 1995 - December 1999 is the resuit of a back calculation. The information collected for the back calculation is not complote but allowed us to properly téike account of the two exceptional revisions occurred in March 2003 and March 2009 for the adoption of the Nace Rcvl.l and the Nace Rcv2 rcspcctivcly.

(25)

to 2005=100. Like GDP, also industrial production is subject to ordinary révisions. Until September 2004, the only ordinary révision of industrial pro

duction data concerned the most recently published observation. Prom Oc- tober 2004 two further révisions were introduced: the first one, conceniing the observations from January to June of the same year, takes place every year iii October; the second one, affecting the préviens three complété years of data, takes place every year in ApriP^. In addition to révisions to raw data, seasonally adjusted data are also subject to the révisions due to the seasonal adjustment procedure. These can dérivé from the adoption of a new seasonal adjustment model^®, or to révisions to raw data. Notice however that season

ally adjusted data are, by construction, subject to révisions every time a new observation is added to the time sériés, even if raw data are not revised, the model has not changed and the parameters are fixed.

1.4.3 The Economie Sentiment Indicator

The ESI is the weighted average of five confidence indicators: the consumer confidence indicator (0.20), the industrial confidence indicator (0.40), the re- tail confidence indicator (0.05), the service confidence indicator (0.30) and the construction confidence indicator (0.05). The real-time data set of the Eco

nomie Sentiment Indicator consists of 119 vintages of time sériés from January 1990: the first one is dated January 31 2001, the last one November 29 2010.

In general, considering vintages of survey data is not necessary as the sea

sonally adjusted time sériés published by the DG ECFIN are not revised^®.

However, there exist vintages of the ESI as the time sériés is re-scaled every year in January in order to hâve long-term average 100 and standard déviation 10. Since real-time data of the ESI are not available from the DG ECFIN, we constructed the real-time data set using the data available on the web at

'^The April révision incorporâtes not only late responses and corrections of eventually inaccurate responses but also the new NA estimâtes released in March. In fact, NA based measures of productivity are used to estimate the industrial production index for those branches in which production is measured by meajis of hours worked.

'®The révision policy of seasonally adjusted data adopted by the ItaJian Institute of Statis- tics for the IPI is the partial concurrent: the models used for seasonal adjustment are tested and eventually changed every year in April and kept fixed until the next March, re-estimating the parameters at each release. The seasonal adjustment model of the IPI has changed in March 2003, March 2009 ajid April 2010.

'®The incthod used by the DGECFIN to clean the data from seasonality, caJled DAIN- TIES, does not imply any révision of past observations: at each release one observation is added to the seasonally adjusted time sériés without changing previously released data. See e.g. Commission (2007, Annex A.2).

(26)

the end of November 2010^®. This allowed us to reproduce the ESIs that were effectively published fiom August 2003 to November 2010. The previous vin

tages do not exactly correspond to the published ones because the latter were based on data classified accordiiig to a different classification of the économie activities, the Nace Rev.l.

The vintages of the ESI are obtained following step by step the method described in Commission (2007, Chapter 3).

1.5 The release calendar of the data and the design of the noweast exercise

The empirical evidence provided in this chapter is the resuit of a genuine real- time out-of-sample exercise. The real-time setting is ensured not only by the use of real-time vintages of published data (see section 1.4), but also by setting up a forecast design in which the noweasts are updated as new information is released taking into account the official release calendar^^.

In Italy a preliminary estimate of GDP, the so called flash estimate, is published about 45 days after the end of the reference quarter. A first revised estimate is then released about 25 days later, together with the whole System of QNA aggregates for the same quarter. To give an example, on February 15 2011 ISTAT released the flash estimate of GDP for the last quarter of 2010;

such estimate was then revised (for the first time) in occasion of the March release (March 10 2011). The IPI is characterized (approximately) by the same publication lag as the flash estimate of GDP (see 1.1), but being recorded at a monthly frequency can provide essential intra-quarter information on GDP.

To give an example, the industrial production referred to the first moiitli of 2010:q4 was released on December 10 2010, about two months earlier than the release of the flash estimate of GDP. Survey data are significantly more timely than the industrial production index: the ESI, in fact, is released at the end of the reference month^^. Then, on December 10, in addition to the IPI registered in October the Economie Sentiment Indicator referred to November

^“Notice that they are potentially affecter! by a break as data from May 2010 are derived from the new classification of the économie activities, the Nace Rev.2. The DG ECFIN published only in February 2011 cohérent long time sériés of B&C survey data according to the new classification.

^'Table 1.1 reports as tm exarnple the exact release ealendar of the data from January to December 2009.

^^Exceptions are the releases referred to December which are published the first working day of January of the following year.

(27)

Instead of following the official calendar exactly, we opted for updating the nowcasts at regular intervals of about fifteen days: in the middle of the month (approximately the 15th day) and at the end of the month (between the 28th and the 31st day). This is sufficient to capture the relevant flow of information (the IPI and GDP are released in the middle of the month and the ESI at the end of the month) and facilitâtes the reading of the results.

Data availability along the quarter according to the adopted release scheme is exemplified in Table 1.2 (first panel) in which white cells indicate missing data, black crosses indicate data availability and red ones indicate just released observations. Moving from left to right we move along the quarter in which we are performing the nowcasts (q), while rows indicate the period to which data are referred to.

1.6 Empirical results

This section illustrâtes the results obtained nowcasting the quarter on quarter growth rate of Italian GDP with the real time-data set described in section 1.4. Particular attention is devoted to the following aspects:

- the évolution of uncertainty of consecutive nowcasts along the quarter - the rôle of information conveyed by survey data

- the compaxison with professional and institutional forecasts.

Our nowcasts axe obtained using the mixed fxequency VAR pxesented in section 1.3 with GDP, the IPI and the ESI. GDP and the IPI enter the model in log différences, while the ESI enters in levels. Augmented Dickey Fuller tests cannot reject the null of unit root in none of the three series^^. Nonetheless we opted for using the ESI in levels because, despite its persistency, being re-scaled every year to hâve constant niean and variance it is by définition a stationary sériés (see Gommission (2007, Chapter 3) for a description of how the ESI is computed). Figure 1.1 shows the time sériés path of GDP, the industrial production index and the sentiment indicator, while Figure 1.2 shows their dynamic cross-correlation. The three variables included in

^^The tests havc been performcd considering the time sériés at their native frequency:

quarterly for GDP and monthly for the IPI and the ESI. Detailed tables of results are available on request.

was available. This means that using monthly information would allow us to hâve information on two months out of three before the end of the quarter.

(28)

the VAR are characterized by common dynamic properties and by a strong cross-correlation. GDP and IPI growth rates are very highly correlated at ail frequencies, confirming the key rôle of industrial production at providing information on GDP. Though slightly lower, the corrélation between the ESI and GDP growth and between the ESI and the IPI growth, more concentrated at low and medium frequencies, supports the use of the ESI to provide early information on both GDP and the IPI.

In addition to the tri-variate mixed-frequency VAR described above (VAR3), we also look at three competitors: a random walk with drift in the levels of logged GDP used as a benchmark (RW), a mixed-frequency VAR with only GDP and the IPI (VAR2) and a bridge équation model with both the IPI and the ESI (BRIDGE). In the bridge équation the missing observations of the monthly indicators at the end of the sample are forecast by making use of auxiliary univariate autoregressive models^^.

Estimâtes of the previous, the current and the following quarter growth rate of GDP are computed every fifteeii days from April 15 2001 to mid of December 2010 on rolling Windows of five years of data^^: the first two estima

tion samples staxt in January 1995 and end in April 2001 (with data available at mid and end of April, respectively), the second two start in January 1995 and end in May 2001, the third two start in January 1995 and end in June 2001; the beginning of the sample is then shifted to the second quarter of 1995: the fourth two samples start in April 1995 and end in July 2001, and so on. The time sériés enter the models according to their historical availability (see the previous section and Table 1.2), so that missing observations are in general présent at the end of the sample.

The évaluation sample goes from the first quarter of 2001 to the third quar

ter of 2010^®. Nowcasts are evaluated in terms of their Mean Square Forecast Error (MSFE). The GDP release against which we evaluate the nowcasts is the flash estimate^^.

^'*The lag spécification of the autoregres.sive models is selected at each updating ro\ind by minimizing the Bayesian Information Criterion (BIC).

^®The results are robust to changes in the length of the estimation sample. We chose to use this length because it provides estimâtes with good properties allowing for a longer évaluation sample.

^®In practice we also hâve nowcasts for the fourth quarter of 2010 and the first quarter of 2011 but they are not used for the évaluation because when data were downloaded GDP for the fourth quarter of 2010 and the first quarter of 2011 were not available.

^’^We also hâve results for alternative target releases of GDP: namely, the first revised estimate and the estimate after one year and after three years from the first revised estimate.

The complété set of results, available upon request, confinns at a qualitative level those reported in the paper for the fiash estimate.

(29)

1.6.1 Main results

The main results are reported in Table 1.3. The évaluation period spans from 2001ql to 2010q3 for forecasts niade about every fifteen days from April 2001 to (mid) December 2010 included. Columns in the Table correspond to the updates made along q of the estimâtes of GDP growth of the previous quarter {q - 1), the current (q) and the next quarter (g + 1). The Table reports the MSFE of the naïve constant growth model (in parenthesis) and the relative MSFE of three alternative models; our mixed-frequency VAR with GDP, the IP index and the ESI (the VAR3), a mixed-frequency VAR without the survey indicator (VAR2) and a bridge équation model exploiting the same informa

tion as the VAR3 (BRIDGE). The relative MSFEs axe computed as ratios between the MSFE of the model (VAR3, VAR2 or BRIDGE) and that of the naïve model, hence numbers smaller than one indicate a better performance with respect to the naïve model. A single(double) asterisk(s) signais a pré

dictive accuracy significantly higher than the benchmark model at the 10(5) per cent level^®.

The table clearly shows that forecast uncertainty decreases as information ar

rives along the quarter. AU of the three models considered outperform the RW at anticipating the estimate of the GDP growth of the previous and the current quarter but the VAR3 is on average affected by a lower forecast un

certainty with respect to the two competitors. In particular, the MSFE of the nowcasts of the previous quarter is about 70 per cent lower than the naïve model, against about 60 per cent for the VAR2 and the bridge équation^®.

The superior performance of the VAR3 with respect to the VAR2 and the BRIDGE is even more évident for the nowcasts of the current quarter GDP.

At this forecast horizon, if fact, its relative MSFE decreases from 0.87 two weeks after the beginning of the quarter to 0.39 the last day of the quarter.

The table also shows that the information on the first month of the quarter reveals fundamental for improving the accuracy of the nowcasts of the current quarter: the release of the ESI at the end of the first month, in fact, reduces the relative MSFE of the VAR3 from 0.87 to 0.68 (the relative MSFE of the VAR2 nowcasts made at the same time is 0.95), while the release of the IPI referred to the first month of the quarter further reduces it from 0.61 to 0.39.

^®Unilatcral Diebold and Mariano tests hâve been performed taking into ciccount the Newey-West correction for licteroscedasticity and autocorrélation in the forecast crrors. De- tailed tables are available on request.

^®The estimâtes of the previous quarter made with the VAR3 at the end of the first month of the quarter do not improve with respect to those made a couple of weeks earlier because the additional information rcleased (ESI) refers to the current quarter and not to the previous one.

(30)

A further resuit emerging from Table 1.3 is that while the VAR2 and the VAR3 significaiitly outperform the naïve model already from the second month of the current quarter, this is not the case for the bridge model, whose MSFE is lower than the RW only from the third month of the quarter^®. This resuit is in line with Golinelli and Parigi (2008) who found that the bridge équations outperform the naïve constant growth model only when full information on the quarter to nowcast is available through the monthly indicators.

When looking at the nowcasts of the next quarter the results show a reverse picture. The forecast uncertainty of the mixed-frequency VAR without the ESI remains comparable with the naïve model while that of the BRIDGE and, especially, of the VAR3 degenerate fast. It seems that the information conveyed by the surveys is crucial for anticipating the GDP release of the previous and of the current quarter, but it is detrimental to forecast at longer horizons^^.

1.6.2 Nowcasting during the économie downturn

The use of intra-quarter information to nowcast GDP has revealed particularly effective during the 2008-2009 économie downturn. The Table 1.4^^ in the Appendix shows that, compared to the naïve model, ail the models perform relatively better during the crisis than over the whole évaluation period, partially due to the severe détérioration of the RW noweasts at turning points.

Our results, though based on a limited number of observations, show that the VAR3 clearly outperforms the other two models. In fact, the VAR2 and the BRIDGE make well at anticipating the GDP quarter-on-quarter growth rate of the previous quarter, but their relative MSFE (0.29 and 0.27, respectively) doubles that of the VAR3 (0.14). The relative MSFE of the VAR3 estimâtes of the current quarter decreases along the quarter from 0.93 to 0.26 during the crisis and only from 0.87 to 0.39 over the full évaluation period. Also the rôle of the information gathered by the survey indicator appears amplified during the crisis: the relative MSFE decreases from 0.93 to 0.68 (from 0.87 to 0.68 over the full sample) thanks to the release of the ESI at the end of the first month of the quarter and from 0.52 to 0.43 (0.65 to 0.61 over the full sample) thanks to the release at the end of the second month of the quarter.

^^Diebold and Mariano tests however cannot reject tlie riull of equ2Ü prédictive accuracy between the RW and the BRIDGE. See Table 1.3.

®’This resuit holds for ail of the survey indicators considered in our prelimiiiary analysis.

See section 1.4.

^^Diebold and Mariano tests of equal prédictive accuracy hâve not been performed due to the small nuinber of observations available.

(31)

A different perspective of the same issue is given by Figure 1.3 that shows the box plot of the nowcasts errors of the four models 2is information cornes available. On the horizontal axis is reported the time in which the nowcasts of GDP q-o-q growth rate of quarter q are updated, staxting from the middle of the first month of the same quarter (mlq) to the end (e) of the first month (1) of the following quarter {q + 1); elq+1. The boxes are limited by the 25th and the 75th percentiles below and above, respectively, while crosses represent nowcast errors outside these bounds. By looking at the picture, it is quite clear how the superiority of the VAR3 mainly consisted of using the increasing amount of information coming from successive data releases more efficiently than the other models, which reduced the amplitude and the occurrence of extreme nowcast errors at each updating round. A comparison with the same box plot referred to the crisis period (see Figure 1.4) confirms the finding that the good performance of the VAR3 is particularly marked during the crisis.

1.6.3 The rôle of survey data

The literature on nowcEisting has recently shown an increasing interest on the rôle of soft data, like business surveys, in nowcasting. The use of soft data is mainly justified by their timeliness with respect the so called hard data, but their effectiveness at reducing forecast uncertainty may be limited by their soft nature.

The practice of introducing information from business surveys into forecasting econometric models has a strong tradition among practitioners, especially in Italy. Production expectations of the manufacturing sector hâve been used to forecast the IPI by Bruno and Lupi (2004), who set up the model used for years by the ISAE for its institutional forecasts. Survey indicators are also key ingrédients in the bridge équation models proposed, among others, by Parigi and Schlitzer (1995) and Golinelli and Parigi (2008) to nowcast GDP growth. Only recently, however, some effort has been made to investi- gate their effective usefulness at improving prédictive accuracy in real-time.

Giannone, Reichlin, and Small (2008) analyzed the informational content of consecutive data releases finding that survey indicators substantially reduce forecast uncertainty because of their timeliness but their informational con

tent diminishes as hard data cornes available. Their results are confirmed by a similar analysis in Bahbura and Rünstler (2007) and Klein and Ozmucur (2010). Differently, Matheson (2010) found evidence that the informational content of survey indicators goes beyond their timeliness.

The design of our nowcast exercise allows us to analyze the real-time infor-