an overall negative relationship is found for other samples (e.g. US-exports and imports,
sector 4). As indicated by the confidence intervals the estimated average local impact of
FX uncertainty on trade growth is mostly insignificant. Conditional on specific levels of
FX uncertainty the slope of the semiparametric estimates differs from the linear estimates.
The latter finding is particularly relevant for scenarios of high FX uncertainty (US exports
in sector 5, imports in sector 4). Locally the linear relationship is not covered by the
semiparametric confidence bands in a few cases (e.g. Imports sectors 0 and 8, exports sector
5). Moreover the latter confidence bands do not uniformly cover the zero line indicating
that although the dependence of trade growth on FX uncertainty is weak over wide ranges
of the conditioning variable it is not uniformly insignificant. In scenarios of relatively low or
high FX uncertainty a systematic impact on trade growth might be present. As pointed out
before, however, the latter result is not uniform over sectors with respect to its sign.
Similar to Figure 1, Figure 2 shows results obtained for sector 7 selected over the cross
section. Again the empirical results do hardly allow any uniform interpretation such that
the relation of interest appears to be sector and country specific and also fails uniformity
when contrasting results for import and export growth. All linear slope estimates θk turn
out to be insignificant at the 5% level. For the majority of semiparametric estimates (e.g.
all export estimates), however, the provided confidence intervals do not uniformly cover the
zero line thereby indicating a locally significant relationship between the two variables.
4 Forecasting
Forecasting is an important area of applied econometrics which provides a complementary
means for model comparison. To uncover the dependence of trade growth on FX uncer-
tainty, however, forecasting exercises have not been used yet. In the spirit of the concept
of Granger causality one would expect some link between the variables of interest if fore-
casts of trade growth conditional on FX volatility improve forecasts obtained when excluding
volatility from the conditioning information set. As a means for model comparison forecast-
ing performance may be preferred over in-sample fitting since factors as the number of model
parameters or the flexibility of the assumed functional relation between dependent and ex-
planatory variables affect this criterion nontrivially. Therefore this section will compare the
scope of linear and nonlinear models in forecasting trade growth conditional on FX uncer-
tainty. Before discussing the forecasting design, however, two further model specifications
used for forecasting are motivated.
12