Michael P. Clements and David F. Hendry*
Some recent developments in the theory of economic forecasting using econometric systems are reviewed. Measures of forecast uncertainty for conditional and unconditional forecasts suggest a limit to forecastability and show the potential advantages of combining types of forecast to improve accuracy; but pooling forecasts from rival econometric models violates encompassing. Mean square forecast error comparisons are criticized in favour of generalized forecast error second moments: only the latter are invariant to linear transformations. We delineate five sources of forecast uncertainty: parameter non-constancy; estimation uncertainty; variable uncertainty; innovation uncertainty; and model misspecification. A theory of intercept corrections to mitigate such errors is discussed. Asymptotic forecast error variance formulae for nonstationary economic time series depend on the treatment of unit roots and cointegration but work well in finite samples, and show that forecast evaluation based on differenced data may fail to reveal inadequate models.
We consider several aspects of the theory of economic forecasting based on econometric models. The term forecast is used to denote a statement about a future event or set of events; prediction is used to denote an implication of a model, so forecasts are a sub-class of predictions. We do not assume that the data generation process (DGP) is constant over time or that the model coincides with that DGP,____________________