observations used to determine the hm-statistic is small (120 in most cases) the respective
number entering the hm-statistics for pooled forecasts is much larger. For instance pooling
over 10 sectors will deliver about 1200 single forecasts. The hm-statistic obtained for all
forecasts is based on 17849 one step ahead predictions. Critical values for the hm-test will
depend on the number of available predictions. For this study we use simulated critical values
taking the exact number of available predictions into account. To determine the latter we
have generated 10000 sequences of bivariate and independent Gaussian random sequences
and used one of these as a forecast for the other.
insert Table V about here
Almost all test statistics vary closely around unity which should not be too surprising
given the low partial degree of explanation reported for particular samples in Figure 1 and
Figure 2. Only a few hm-statistics indicate significantly successful forecasting models. To
facilitate the interpretation of the results bold entries in Table V indicate hm-statistics
exceeding unity with 5% significance. In a few cases of single data sets (specific on country
and sector) nonlinear forecasts of export growth outperform linear forecasts significantly
(e.g. export forecasts for BE or CA, sector 0, NO, sector 3) whereas the opposite pattern of
a significant hm-statistic obtained from linear modelling and insignificant hm-statistics for
nonlinear forecasts is observed only once (NO, sector 5). For a few data sets (SW, sectors 1
and 7, GE sector 4) the hm-statistics are about 1.3 and, thus, show that for these samples
FX uncertainty is helpful in determining the future direction of export growth. Regarding
the pool of all export growth forecasts the threshold specification delivers a hm-statistic of
1.02 which is small but owing to the huge number of forecasts significant at the 5% level.
Both competitors, the linear and semiparametric forecasts, show insignificant hm-statistics
on the pooled level. As already seen for the estimation results (Figure 1 and Figure 2)
the pattern of hm-statistics is again not uniform neither over sectors nor over countries.
Pooling over all sectors in particular the direction of Swedish and Norwegian trade figures
can be detected conditional on FX uncertainty. The latter finding is robust over the entire
support of the conditioning variable. Focussing the attention on scenarios of higher and
lower FX uncertainty it turns out that when pooling over countries the direction of trade
can be determined conditional on volatility in sectors 1, 8, and 7 with the latter being the
most active sector of international trade flows. Recall, however, that all findings obtained on
pooled levels do not uniformly hold over sectors or countries. With respect to forecasting the
direction of import growth conditional on FX uncertainty it turns out that hm-statistics are
somewhat smaller on average and significantly exceed unity only in a few cases. Sector and
country specific hm-test statistics are not given here to economize on space but are available
from the authors upon request.
insert Table VI about here
16