Real‐Time Nowcasting of GDP: A Factor Model vs. Professional Forecasters
| Published date | 01 December 2014 |
| Date | 01 December 2014 |
| DOI | http://doi.org/10.1111/obes.12047 |
| Author | Joelle Liebermann |
783
©2013 The Department of Economics, University of Oxford and JohnWiley & Sons Ltd.
OXFORD BULLETIN OF ECONOMICSAND STATISTICS, 76, 6 (2014) 0305–9049
doi: 10.1111/obes.12047
Real-Time Nowcasting of GDP:A Factor Model vs.
Professional Forecasters
Joelle Liebermann
Central Bank of Ireland, Dame Street PO Box 559, Dublin 2 Ireland
(e-mail: jlieberm@ulb.ac.be)
Abstract
We perform a fully real-time nowcasting (forecasting) exercise of US GDP growth using
Giannone et al.’s (2008) factor model framework. To this end, we have constructed a
real-time database of vintages from 1997 to 2010 for a panel of variables, enabling us
to reproduce, for any given day in that range, the exact information that was available to
a real-time forecaster. We track the daily evolution of the model performance along the
real-time data flow and find that the precision of the nowcasts increases with information
releases and the model fares well relative to the Survey of Professional Forecasters (SPF).
I. Introduction
Assessing the state of the economy in real-time, that is, current quarter real GDP growth, is
of paramount importance to policy-makers, financial market participants and businesses.
For the US, the first estimate of GDP is released only about one month after the end
of the reference quarter. Meanwhile, higher frequency conjunctural indicators releases,
which convey within quarter information, can be used to produce a timely nowcast of
current quarter growth. The information available in the numerous monthly variables can
be summarized efficiently by a small number of common factors, and hence overcome
the curse of dimensionality problem. Factor models, by taking into account information
on many predictors, provide more accurate forecasts of macroeconomic variables than
standard econometric benchmarks (see Stock and Watson,2002a, b; Bernanke and Boivin,
2003; Boivin and Ng, 2005; Forni et al., 2005; Giannone et al. 2008; D’Agostino and
Giannone, 2012, among others). However, in real-time, one has to deal with a further issue
as variables are released on different dates and with varying degrees of publication lags.
Non-synchronous releases of data results in an unbalanced panel at the end of the sample,
that is, a ‘jagged’ edge structure.
Giannone et al.’s (2008) (GRS) developed a framework for real-time nowcasting (and
forecasting) that can handle a large and unbalanced data set. Their model is an automatic,
judgment free procedure that can be updated at any time with information releases, and
JEL Classification numbers: E52, C53, C33
784 Bulletin
hence provides a timely and up-to-date estimation of the state of the economy.They assess
their model performance at nowcasting US GDP growth over the period 1995–2004 using
a pseudo real-time setting. That is, given a panel of revised data as of March 2005, they
replicate the pattern of data availability by aggregating the variables in a stylized monthly
calendar of 15 releases which is kept constant overthe sample. They find that these releases
provide relevant information for nowcasting GDP and that the model performs as well as
the SPF. Other studies have applied the framework of GRS for short-term forecasting of
GDP in a pseudo real-time setting: Ba´nbura and R¨unstler (2011) and Angelini et al. (2011)
among others for the Euro area, Aastveit andTrovik (2012) for Norway, D’Agostino, et al.
(2012) and Liebermann (2012) for Ireland and Marcellino and Schumacher (2010) for
Germany. Ba´nbura and Modugno (2010) and Ba´nbura, et al. (2011) further provide some
extensions to GRS model and apply it to the Euro area.
In this article, we perform a fully real-time nowcasting (and forecasting) exerciseof US
real GDP growth using GRS’s model.To this end, we haveconstructed a real-time database1
of vintages for every day from 1 January, 1997 to 30 June, 2010, for a large panel of US
macroeconomic series. This is important as data available and used by forecasters and
policy-makers in a real-time setting are preliminary and differ from ex-post revised data,
and the order of releases for some variables is not constant. Given that data revisions may
be quite substantial, the use of revised instead of real-time data may not be innocuous for
forecasting. Faust and Wright (2009), for example, argue that the practical relevance for
forecasting of findings based on revised data is on open issue. Croushore (2011)2provides
an extensive survey of the impact of using latest-available, that is, revised, data instead of
real-time on empirical results and shows that it can substantially affect findings.
Our database of vintages3includes monthly series like soft data (surveys), interest rates
(term and credit spreads) and hard data such as industrial production, employment, retail
sales, housing, income and spending and prices among others. Hence, it coversa wide range
of conjunctural indicators which are typically used to assess the state of the economy. This
last point was emphasized by Bernanke and Boivin (2003) who found that for forecasting
performance it is important to have a ‘data-rich’panel in the sense that it covers a wide scope
of indicators. These vintages enable us to reproduce the exact information available to a
real-time forecaster on any given day over our sample range. Importantly, we can compute
model based nowcasts (forecasts) matching the data available to the SPF participants as of
the survey deadline date, and hence run a realistic nowcasting (forecasting) horse race.4
Moreover, our extended sample (which ends in June 2010) compared to the previously
mentioned nowcasting studies allows us also to examine how the model performed over
the recent recession.
1For most of the series (around 70%) real-time information was gathered from the Federal Reserve Bank of St
Louis ALFRED database (see Appendix A).
2See also Croushore and Stark (2002), Stark and Croushore (2002) and Bernanke and Boivin (2003) on this issue.
3Note that our database of vintages differs from the Federal Reserve Bank of Philadelphia real-time database as
in the latter case vintages for a panel of variables are constructed only once per quarter, around the middle of the
quarter. Whereas in our case, the vintages are availablefor every day in a month, and moreover cover a wider scope
of monthly indicators.
4Note that Stark (2010) also highlighted this point and evaluates the SPF against simple univariate benchmarks
estimated in real-time. He finds that the SPF forecasting performance for GDP growth deteriorates as the actuals used
for forecasts evaluation are revised, but that data revisions do not affectthe relative performance of the SPF much.
©2013 The Department of Economics, University of Oxford and JohnWiley & Sons Ltd
To continue reading
Request your trial