Page 373 - Special Topic Session (STS) - Volume 3

STS550 Kyle Hood et al.
factor number selection, move on to a discussion of revisions by PCE services
group, and finally discuss revisions by algorithm type.
Because of the number of parameters to be estimated in the richest lag
structures, we set an upper limit of 2 common factors. The Bai-Ng criterion
selects 2 as the number of factors, so in all cases, 2 factors are considered.
Table 2 provides a summary of the estimation results. This summary is focused
on revision reductions for the individual PCE component series, grouped by
algorithm and PCE services grouping. Columns (2) through (9) represent
counts of detailed components for which the algorithm performed the best.
Columns (3) through (7) show the number of times a model-averaging
algorithm was optimal, and columns (8) and (9) show counts of components
for which single indicator (GB) and factor (BF) models chosen by AICC were
optimal.
For 16 of the 85 series that were nowcasted, no improvement was seen in
the out-of-sample validation set relative to the current method. In the
remaining 69 series, at least one of the methods showed some improvement
over current methods. Improvements were relatively evenly spaced out over
the PCE services component groups. Healthcare showed improvement in 15
of 20 series, recreation in 14 of 16 series, communications in 4 of 6 series,
professional services in all 5 series, travel and transportation in all 14 series,
personal services in 5 of 7 series, and social services in 12 of 17 series. The
column on the far right shows the average relative revision (in root mean
square terms), compared to the current method. The best improvement in a
components series was in motor vehicle rental (in the TRSFTR group), which
showed a revision reduction of about 75%. Average reductions by grouping
ranged from about half (PRS) to 12.6% (RCA), with an overall average of 26.8%.
As noted above, the forecast combination puzzle asserts that empirically,
simple averages often perform better than other methods that derive weights
from model performance (Smith and Wallis, 2009). This holds for these results.
For 34 of the series (nearly half of the 69 for which a reduction in revision was
achieved), simple averages (means or medians) of the 12 forecasts were the
best-performing algorithms. For 26 series, the best model was a single model
within either the indicator or factor models selected by AICC. Among these,
the indicator models was selected more than three times as often as the factor
model. In only 9 cases were the other 3 methods optimal, with the AIC-based
and the Bates-Granger methods accounting for 8 of the 9.

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