Combining Model‐Based Near‐Term GDP Forecasts and Judgmental Forecasts: A Real‐Time Exercise for the G7 Countries
| Author | Jasper M. de Winter,W. Jos Jansen |
| Date | 01 December 2018 |
| Published date | 01 December 2018 |
| DOI | http://doi.org/10.1111/obes.12250 |
1213
©2018 The Department of Economics, University of Oxford and JohnWiley & Sons Ltd.
OXFORD BULLETIN OF ECONOMICSAND STATISTICS, 80, 6 (2018) 0305–9049
doi: 10.1111/obes.12250
Combining Model-Based Near-Term GDP Forecasts
and Judgmental Forecasts:A Real-Time Exercise for
the G7 Countries
W. Jos Jansen† and Jasper M. de Winter‡
†Financial and Economic Policy Department, Ministry of Finance, The Hague, The
Netherlands (e-mail: w.j.jansen@minfin.nl)
‡Economics and Research Division, De Nederlandsche Bank, Amsterdam, The Netherlands
(e-mail: j.m.de.winter@dnb.nl)
Abstract
We investigate the effects of combining model-based near-term GDP forecasts and judg-
mental (quarterly) forecasts by professional analysts (Consensus forecasts) in a real-time
setting for the G7 countries over the years1999–2013. Model-based forecasts are produced
by a dynamic factor model (DFM). We consider as combination schemes the weighted
average and the linear combination. Combining with subjective information delivers siz-
able gains in forecasting ability of statistical models for all countries except Japan, even
when subjective forecasts are somewhatdated. Accuracy gains are much more pronounced
in the volatile period after 2008 due to a marked improvement in predictive power of Con-
sensus forecasts relative to the DFM. A possible explanation is that mechanical models
may be more vulnerable to extreme observations in estimation samples. Consensus fore-
casts are superior at the moment of publication for most countries since 2008. For some
countries forecast combinations can improve upon Consensus forecasts in between the
quarterly release dates of the Consensus survey.
I. Introduction
Policy makers and economic agents have to make decisions in real time on the basis
of incomplete and inaccurate information on current economic conditions. For example,
data on real GDP, which is the broadest measure of aggregate economic activity, are
released on a quarterly basis with a substantial time lag (six weeks in many advanced
countries) and are subject to revisions. However, a wealth of statistical information that is
JEL Classification numbers: C33, C53, E37.
*The opinions expressed in this paper are the authors’ personal views and do not necessarilyreflect the position
of De Nederlandsche Bank or the Ministry of Finance. Weare grateful to Editor Francesco Zanetti, two anonymous
referees and seminar, workshop and conference participants at De Nederlandsche Bank, the Deutsche Bundesbank,
the International Symposium on Forecasting 2015 in Riverside, the Computational and Financial Econometrics
Conference 2016 in Seville and the Society for Economic Measurement Conference 2017 in Cambridge for their
helpful comments. Weare particularly indebted to Job Swank for his valuable comments and stimulating suggestions.
1214 Bulletin
directly and indirectly related to economic activity is nowadays available from public and
private sources. Policymakers, firms and financial market participants may exploit this vast
body of statistical information to form expectations on the current state of the economy
and its near-term development. This requires solving the practical problem of handling
a large-scale information set of potentially hundreds of time series that are observed at
different frequencies and with different publication lags (the so-called ragged edge prob-
lem). The recent nowcasting literature has developed several statistical methodologies for
generating near-term GDP forecasts based on large mixed-frequency data sets with ragged
edges. Examples are bridge models, factor models, mixed-data sampling models (MI-
DAS), mixed-frequency vector-autoregressive (MFVAR) models and Bayesian VARs and
mixed-frequency models.1
Apart from model-based forecasts, policy makers and economic agents may also take
advantage of published forecasts made by professional analysts. From a practical point of
view, such forecasts are cheap and easy to use. Currently,several surveys on the economic
outlook are available on a regular basis. The Federal Reserve Bank of Philadelphia and
the European Central Bank both maintain a regular Survey of Professional Forecasters.
Moreover, the survey firm Consensus Economics publishes a well-known compilation of
macroeconomic forecasts by professional forecasters for many countries. Model-based
forecasts are the result of purely mechanical recipes using statistical data and do not in-
corporate subjective elements. By contrast, forecasts by professional analysts reflect much
more information than statistical data, which are inevitably limited. Surveys among pro-
fessional forecasters by European Central Bank (2009, 2014), Stark (2013) and D¨opke,
Fritsche and Waldhof (2017) indicate that GDP forecasts mostly reflect a mixture of model-
based and judgmental information. For example, 60% of the respondents in the European
Central Bank (2014) survey consider their short-run forecasts as being model-based but
with judgmental adjustments, while 24% think they are essentially judgment-based (expe-
rience and intuition) and 16% purely model-based. In Stark (2013), 80% of the panelists
view their projection method as a statistical model with subjective adjustments.
The evidence in Aiolfi, Capistr´an and Timmermann (2011), Jansen et al. (2016) and
Liebermann (2014) suggests that subjective forecasts by private sector analysts often em-
body valuable information that sophisticated mechanical forecasting procedures fail to pick
up. Publicly available subjective forecasts thus seem to offer the potential of enhancing
real-time model-based GDP forecasts and thus a better assessment of the current state of
the economy. The main purpose of this paper is to investigate whetherforecasts by analysts
are actually able to improve GDP forecasts generated by purely statistical procedures in
real time. We also pay attention to the reverse question whethermodel-based forecasts can
be used to enhance subjective forecasts. Our procedure takes into account the information
availability constraints facing practitioners when running forecasting models and forming
expectations. As our benchmark, we use forecasts produced by a dynamic factor model
estimated on real-time data vintages using a rolling estimation window of 15 years. We
employ the quarterly forecasts published by Consensus Economics as our measure of
1See among others Baffigi, Golinelli and Parigi (2004), Angelini et al. (2011), Stock andWatson (2011), Kuzin,
Marcellino and Schumacher (2011), Ghysels, Sinko and Valkanov (2007), Marcellino and Schumacher (2010),
Ba´nbura, Giannone and Reichlin (2010), Carriero, Clark and Marcellino (2015) and Jansen, Jin and de Winter
(2016).
©2018 The Department of Economics, University of Oxford and JohnWiley & Sons Ltd
To continue reading
Request your trial