Abstract/Summary

After reviewing the history of analyses of economic forecasting, the role of econometrics in improving economic forecasting is considered, building on CLEMENTS and HENDRY (1994a). The basis of the analysis is a world where model selection is difficult, no model coincides with the economic mechanism, and that mechanism is both non-stationary and evolves over time. On the constructive side, econometric analysis suggests ways of reducing each of the resulting five sources of forecast uncertainty (parameter non-constancy; estimation uncertainty; variable uncertainty; innovation uncertainty; and model mis-specification). On the critical side, the lack of invariance of forecast evaluation procedures to the representation of the model may camouflage inadequate models. We show that forecasts generated from vector autoregressions in differences may be more robust to certain forms of structural change over the forecast period, and that a similar result can be achieved by suitable forms of intercept corrections in vector error-correction mechanisms.