Page 368 - Special Topic Session (STS) - Volume 3
P. 368
STS550 Kyle Hood et al.
(consumption, investment, government expenditures, exports and imports),
with personal consumption expenditures on services (PCE services) showing
particularly large revisions due to lack of availability of source data for the first
estimate in most detailed components. In this paper, we focus on these detailed
components, using a combination of nowcasting and model-averaging
techniques to reduce the size of revisions.
A lack of source data for the first estimate of most PCE services
components means that Bureau of Economic Analysis (BEA) must use other
indicators for these components in the first estimate. These indicators are
chosen to match as well as possible the component being considered. For
some estimates, the indicator is growth in population and a component-
specific price index, while for other estimates, the indicator is wage and salary
growth in the industry associated with the component. Unfortunately, these
indicators are often weakly or even negatively correlated with the third-
estimate values of components which use the Quarterly Services Survey (QSS)
as an indicator for the third estimate. For example, for health care services,
correlations between the indicator and third estimate are negative for 4 of 15
detailed components and are above 0.5 in only 2 of 15 cases. Similar patterns
hold for other categories.
Lack of a close correlation between indicators and the target series can be
alleviated by the addition of other useful data. The indicators used for the first
estimate do not incorporate additional information that is available at the time
of the first estimate. For example, medium-term movements in the target series
could be relevant, as well as long-term trends. In addition, information from
more general movements in PCE services could be relevant, but the current
method restricts indicators to those matching the specific components.
Because in this paper we are interested in estimating the recent past where
data have not become available yet, we exploit the nowcasting literature.
Nowcasting is defined as the prediction of the present, the very near future
and the very recent past (Giannone, Reichlin and Small, 2008). It is different
from forecasting in some specific data-related problems that must be
overcome. In this case, we have access to monthly indicators for our quarterly
series and we must deal with the fact that our indicators are preliminary data
which are likely to be subsequently revised. Other issues specific to nowcasting
such as the “ragged edge” problem do not arise (cf. Giannone, Reichlin and
Small, 2008). For a survey of relevant techniques, see Foroni and Marcellino
(2013).
In this paper, we expand work from an earlier paper (Chen and Hood, 2018)
in which we showed that two nowcasting techniques, the General Bridge
Equation (GB) (Klein and Sojo, 1989) and Bridging with Factors (BF) models
(Giannone, Reichlin and Small, 2008), reduce revisions for many PCE services
detailed components. Here, we combine these two methods (in an array of
357 | I S I W S C 2 0 1 9