82nd International Atlantic Economic Conference

October 13 - 16, 2016 | Washington, USA

Evaluating the Bloomberg survey forecasters

Saturday, October 15, 2016: 2:55 PM
Constantin Burgi, MSc , Economics, The George Washington University, Washington, DC
In this paper, we look at the quarterly Bloomberg survey of forecasters collected since 1993 for real GDP growth, the unemployment rate, CPI inflation and 10-year government bond yields. The survey collects forecasts for the current quarter as well as up to 4 quarters ahead. This survey is only available within Bloomberg terminals with the command ECFC. (Bloomberg Brief, June 2014, snapshot of the average annual forecasts in the survey at that time: http://www.bloombergbriefs.com/content/uploads/sites/2/2014/07/G20_Economic_Forecasts_June2014.pdf) We evaluate the forecasts collected by this survey by looking at overall biasedness of the forecasts, by checking if forecast revisions are efficient, by comparing the forecast performance of the survey relative to the random walk and by testing if there are better and worse forecasters. We find that forecasts are largely unbiased without controlling for the state of the economy and their revisions are efficient for GDP but not for bond yields or the unemployment rate. There also are better and worse forecasters ex post for all variables and horizons, but only for bond yields, they can be detected ex ante. This allows us to create a subset of forecasters that can beat the simple average statistically significantly. After this evaluation, we look at the distribution of forecasters, to check, if this provides additional information. We initially check, how often there is a consensus based and find that this is largly the case for GDP and bond yields, but not for the unemployment rate. We then use the spread of forecasters as measured through the standard deviation to predict recessions. We find that the performance of our indicator improves significantly due to the addition of this measure in sample, even though the sample size is small.