EFFECTS OF THE INCORRECT IDENTIFICATION OF NON-STATIONARITY IN MEAN AND IN VARIANCE FORECASTING OF ECONOMETRIC MODELS

The purpose of the paper was to evaluate the effects of the incorrect identification of non-stationarity in mean and in variance for the forecasting behaviour of econometric models specified for levels (the TS model - strategy 'always take levels') and for differences (the model DS and EC - strategy 'always difference'). To evaluate these effeets the Monte Carlo experiments were carried out for data generated under the assumption of the same and different dependence in the whole frequency band. The comparison of the strategy 'always take levels' and the strategy 'always difference' indicated that none of these approaches dominates in forecasting. In other words, the models for levels (the TS models) and models for differences (the EC models) can rival even in the case of the incorrect identification of non-stationarity in mean and in variance. These results suggest the usefulness of models for levels as well as models for differences independentiy of the type of non-stationarity (in mean or in variance) in practical application, but provided that these models satisfy the congruence postulate consisting in specifying the model for both the levels and the differences in such a way that the residual process has white noise properties.