SAMPLE SELECTION PROBLEMS IN A MACROECONOMETRIC MODEL CONTEXT -- SOME FURTHER RESULTS

The selection of sample size and time period is vital to econometric modelling since both affect model behaviour and results. Despite this importance, the question has attracted little attention in the literature.In an earlier paper, we analysed the relationship between forecasting performance and sample length for the medium-sized macroeconometric RWI-business cycle model. The study was conducted primarily on the single-equation level examining 1-, 4-, and 8-quarter forecasts for each equation. The "moving-window" size ranged from 20 to 60 in steps of 10 quarters on the longest possible sets of data with the same specification for each sample and window. In general, the results did not reject the prevailing practice of basing estimations on a uniform window size, covering the last 40 quarters of the data base ("moving window"). Both shorter and longer window sizes had advantages for half of the equations, but the improvements in forecasting accuracy in these cases were not very impressive.This paper extends these earlier results by exploiting the information gained from them and by paying attention to the time profile. First, we relax the fixed specification through all samples for particular equations. The 20-quarter window seemed best for the Government sector, not surprisingly, since institutional changes (e.g., of the tax code) are frequent here. These changes are usually incorporated in the equations via dummies and, when forecasting, by add factoring. In capturing such information, the optimal window size will certainly change -- probably also providing interesting information about the effectiveness of the add factoring. Second, the time profile of the forecasts as well as of the quality of estimation of the model blocks is analysed. These may lead to some hints for general structural shifts in the economy.