Page 374 - Special Topic Session (STS) - Volume 3
P. 374
STS550 Kyle Hood et al.
Table 2. Counts of best model average or selection by PCE services group
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11)
PCE No IC-based Bates- Simple averaging Model Average
3
averaging Granger selection
1
2
Group Improvement AIC BIC averaging Equal weights Median Indicator Factor Total Relative
RMSR
HLC 5 1 0 1 1 6 6 0 20 67.3%
RCA 2 1 0 2 2 1 7 1 16 87.4%
COM 2 0 0 0 1 2 0 1 6 74.2%
PRS 0 1 0 1 1 0 2 0 5 48.3%
TRSFTR 0 1 1 0 5 4 2 1 14 73.4%
PER 2 0 0 0 1 2 2 0 7 74.7%
SOC 5 0 0 0 4 4 1 3 17 73.3%
Overall
Total 16 4 1 4 15 19 20 6 85
average
Percentage 18.8% 4.7% 1.2% 4.7% 17.6% 22.4% 23.5% 7.1% 100.0% 73.20%
Notes:
1. HLC: Health care, RCA: Recreation, COM: Communications, PRS: Professional services,
TRSFTR: Travel/transportation, PER: Personal services, SOC: Social services
2. Unweighted totals by component within each group (including only components with
differing indicators for first and third estimate)
3. Unweighted mean
4. Conclusion
Model averaging has the potential to be a powerful tool in reducing
revisions in national economic accounts statistics at the detailed-component
level. For some detailed components, these techniques can reduce revisions
by half or more. However, because of short time series, the forecast
combination puzzle is especially relevant. Simple averages of the models
under consideration performed best nearly half of the time. Nevertheless, it
does appear that other model-averaging techniques may perform well under
certain circumstances. Furthermore, the GB model seems to frequently
outperform the BF model when a single model is optimal, suggesting that
many of the indicators already used to produce these PCE component
estimates contain relevant and useful information on the movements of the
underlying series. Over all, it is clear that there is not broad agreement on a
single approach or algorithm that can be applied across all of these
components. As such, a hybrid approach that uses a specific algorithm for each
detailed component may offer the best option.
References
1. Bai, J. and S. Ng, “Determining the number of factors in approximate
factor models,” Econometrica, 70(1): 191–221.
363 | I S I W S C 2 0 1 9