Econometric forecasting is alive and well. In fact, forecasting is re-emerging as an exciting and vital research area, fuelled not only by its tremendous practical importance, as always, but also by recent advances in both analytic methods (e.g. local-to-unity and long-horizon asymptotics) and computational methods (e.g. simulation techniques). The new methods and models, however, are very different from those of twenty-five years ago. This issue contains seven papers that report on various aspects of the new research. The impetus for the issue was a conference on economic forecasting held at the National Bureau of Economic Research (NBER) in Cambridge, Massachusetts on 22 April 1995, jointly sponsored by the NBER, the National Science Foundation and the Journal of Applied Econometrics. Several of the papers were discussed at the conference.
Granger begins the issue with a provocative article on how we might improve both the actual and the perceived quality of economic forecasts. To improve actual forecasts, he argues that more attention should be paid to including the correct information in the conditioning set, and he shows how past forecast errors can help eliminate bias. To improve the perceived quality of forecasts Granger suggests that we communicate forecast uncertainty using 50% confidence intervals; these are much narrower than 95% intervals and may be more useful to decision makers. Finally, Granger argues that structural instability is the most important problem facing forecasters, and he predicts that it will be a major focus of future research.
Clements and Hendry, consistent with Granger's prediction, examine the effects of structural change on the forecasting performance of different models, including a vector error correction model (VECM), a VAR in first differences (which ignores cointegration), and a model that incorporates 'intercept corrections'. They provide analytic results on the forecasting performance of the various models when the coefficients of deterministic components (constants and time trends) have discrete breaks, under the assumption that the parameters are otherwise known. In that case, differenced VARs are robust to specific forms of structural change (a change in the mean of the error-correction term) and in such cases provide more accurate forecasts. Clements and Hendry also use the three models to forecast UK wages, prices and unemployment; they find that the accuracy of forecasts from the differenced VAR compares favourably to that of forecasts from the VECM.