How far can we forecast? Statistical tests of the predictive content

We develop tests for the null hypothesis that forecasts become uninformative beyond some maximum forecast horizon h*. The forecast may result from a survey of forecasters or from an estimated parametric model. The first class of tests compares the mean-squared prediction error of the forecast to the variance of the evaluation sample, whereas the second class of tests compares it with the mean-squared prediction error of the recursive mean. We show that the forecast comparison may easily be performed by adopting the encompassing principle, which results in simple regression tests with standard asymptotic inference. Our tests are applied to forecasts of macroeconomic key variables from the survey of Consensus Economics. The results suggest that these forecasts are barely informative beyond two to four quarters ahead