Error-Correction Versus Differencing in Macroeconometric Forecasting

Recent work by Clements and Hendry have shown why forecasting systems that are in terms of differences, dVARs, can be more accurate than economet- ric models that include levels variables, ECMs. For example, dVAR forecasts are insulated from parameter non-constancies in the long run mean of the cointegration relationships. In this paper, the practical relevance of these is- sues are investigated for RIMINI, the quarterly macroeconometric model used in Norges Bank (The Central Bank of Norway), which we take as an example of an ECM forecasting model. We develop two dVAR versions of the full RIMINI model and compare ECM and dVAR forecasts for the period 1992.1- 1994.4. In addition we compare forecasts from the full scale models with those of univariate dVAR type models. The results seems to confirm the relevance of several important theoretical insights. dVAR forecasts appear to provide some immunity against parameter non-constancies that could seriously bias the ECM forecasts. On the other hand however, for open systems like the RI- MINI model, the misspecification resulting from omitting levels information seems to generate substantial biases in the dVAR forecasts. Therefore, the incumbent ECM performs comparatively well over the forecast period inves- tigated in this paper.