Evaluating Forecast Uncertainty

Abstract

The National Institute has been producing forecasts since 1958, and they are an essential part of our contribution to the policy debate in the UK and elsewhere. Model-based forecasts for the UK have been published every quarter1 since the 1970s, whilst model-based forecasts for the US and the major Europeans started only in the late 1980s. The country coverage of the European forecasts has increased over the years, and Euro Area forecasts began as soon as membership was clear. There is a long enough history of forecasting to be able to evaluate our performance, and this note extends the work reported in Pain, Riley and Weale (2001), Barrell, Kirby and Metz (2005) and Barrell and Metz (2006). The Institute forecasts for the UK are the main focus of the analysis in this note, but evaluations of the US and Euro Area forecasts produced by the Institute are also undertaken. We focus on evaluating the last forecast undertaken (at T–1) before the start of the year being forecast (at T), as they contain as much information as was possible without using any from the year being forecast. We also look at the evolution of forecasts over the two years preceding the publication of the first outturn for the year being forecast in order to gauge at which point those forecasts become efficient in that the root mean squared deviation (RMSD) of the errors is smaller than standard deviation of the series.