Page 303 - Special Topic Session (STS) - Volume 3

STS544 M. Camachoa et al.
Given the publication lags of GDP, different nowcasting and forecasting
methods have been proposed to asses current and future economic
conditions, using real-time data of economic variables that are published on
a more frequent and timely basis, such as industrial production, financial
variables, or consumer and business confidence. To deal with the mixed
frequencies, unbalanced panel data and the potential non-linearity of these
variables, different methods have been proposed, such as bridge equations,
MIxed DAta Sampling (MIDAS) regression models, and Dynamic Factor
1
Models (DFM).
In this paper, we review the experience at BBVA Research in the use of
DFM to nowcast and forecast GDP growth in our footprint, which can be
summarized in six lessons. First, DFM forecast GDP growth and recession
probabilities at least as well as other models in the large sample of countries
(world, USA, EMU, China, Spain, Portugal, Turkey, Argentina and Mexico).
Second, DFM forecast in a very parsimonious ways, allowing to present easily
the contribution of different indicators to forecasts innovations. Third, financial
variables, such as the slope of the yield curve or financial tension indexes,
contain valuable information about future growth and can be easily
introduced in DFM. Fourth, DFM should be tailored to different countries and
variables. Fifth, DFM can be used to estimate the underlying activity in
countries where official GDP statistics are not reliable. And sixth, DFM allow
the introduction of useful indicators of economic activity obtained using real-
time big data (e.g., retail sales, credit cards spending, etc.), improving
nowcasting significantly.

2. Methodology
2.1 Small-scale dynamic factor models
DFM were advocated by Geweke (1977) as a time-series extension of factor
models previously developed for cross-sectional data. The premise of DFM is
that the covariation among economic time series variables at leads and lags
can be traced to a few underlying unobserved series, usually known as factors.
Although dynamic factor models have been the source of a vast literature, in
this paper we focus on the "single-index" dynamic factor model developed by
Stock and Watson (1989, 1991). Low-dimensional parametric dynamic factor
models are expressed in state space form. This implies that the Kalman Filter
can be used to construct the Gaussian likelihood function and thereby to
estimate the unknown parameters of the model by maximum likelihood.
2.2 The single-index dynamic factor model
Let denote the × 1 vector of a set of macroeconomic indicators

observed at period . These indicators are assumed to be covariance stationary

1 See Camacho, Perez-Quiros and Saiz, 2013, and Foroni and Marcellino, 2013.

292 | I S I W S C 2 0 1 9