Forecasting Substantial Data Revisions in the Presence of Model Uncertainty*


  • *

    We thank Dean Croushore, James Mitchell, Athanasios Orphanides, Adrian Pagan, Simon Potter, Heather Robinson, Andrew Scott, Norman Swanson, Simon van Norden, James Yetman and two referees for helpful comments. We are also grateful to seminar participants at the Society for Computational Economics 2005 meetings, the CIRANO Data Revisions Workshop, University of New South Wales, University of Otago, RBNZ, Norges Bank, FRB San Francisco and the North American Summer Econometric Society Meetings 2006. Financial support from the ESRC (Research Grant No RES-000-22-1342) is acknowledged gratefully. The views in this article do not reflect those of the Reserve Bank of New Zealand or Norges Bank.


A recent revision to the preliminary measurement of GDP(E) growth for 2003Q2 caused considerable press attention, provoked a public enquiry and prompted a number of reforms to UK statistical reporting procedures. In this article, we compute the probability of ‘substantial revisions’ that are greater (in absolute value) than the controversial 2003 revision. The predictive densities are derived from Bayesian model averaging over a wide set of forecasting models including linear, structural break and regime-switching models with and without heteroscedasticity. Ignoring the nonlinearities and model uncertainty yields misleading predictives and obscures recent improvements in the quality of preliminary UK macroeconomic measurements.