linear model will be suitable to provide best linear forecasts conditional on the data. In
this case the average rank of such forecasts should be smaller than 2.5. Accordingly, a case
where nonlinear models as the threshold or semiparametric forecasting rules deliver average
rank statistics smaller than the respective outcome of the linear forecasting scheme could
be interpreted as evidence in favor of a nonlinear relation linking FX uncertainty and trade
growth.
4.3 Results
Forecast error correlation: The right hand side panels of Table IV show absolute rejec-
tion frequencies obtained from LM-tests against serial correlation of one step ahead forecast
errors at the 5% significance level. Similar to ex-post diagnostics the results are given in
form of aggregates over 10 economic sectors. With respect to autocorrelation of ex-ante
forecast errors no striking differences are obtained when comparing the outcomes of fore-
casting export growth on the one hand and import growth on the other. First order error
correlation is diagnosed in about 15.5% of all employed forecasting models which exceeds
the nominal significance level by far. When testing for serial correlation up to order 12 we
again encounter the problem that seasonal patterns of trade growth may not be entirely
captured by the selected single equation models. The corresponding test statistic (AR12)
yields a rejection of the hypothesis of uninformative one step ahead forecast errors for about
41.2% of all empirical models. Owing to heteroskedasticity and the presence of a few huge
outliers in most sequences of forecast errors, however, we do not interpret the latter find-
ings as indicating severe misspecification of the employed forecasting schemes. Comparing
alternative forecasting procedures it turns out that on average the semiparametric model
yields error sequences which show serial dependence less often. For example with respect to
first order testing the latter procedure gives 19 rejections of the respective null hypothesis
when forecasting import or export growth whereas the linear regression shows significant
autocorrelation for 25 and 23 error sequences when forecasting export and import growth,
respectively. For both, export and import models, 24 error sequences obtained from ”zero”
forecasts show significant first order autocorrelation.
Henrikkson Merton tests: Table V reports country and sector specific hm-test statistics
characterizing alternative forecasting schemes for export growth. Moreover, hm-statistics are
given for country and sector specific pooled predictions and for an aggregate obtained over
all one step ahead forecasts of export growth. Since the relative forecasting performance
of alternative procedures may be different over states of lower and higher FX uncertainty
relative to states of medium FX uncertainty we also provide hm-statistics for pooled forecasts
where the conditioning variable vtk is outside its interquartile range. Similar to the hm-
statistics obtained for pooled forecasts of export growth Table V shows analogous results for
pooled forecasts of import growth. Note that whereas for each single data set the number of
15