Abstract

Policymakers cross-check their projections for multiple variables and forecast horizons
with experts’ forecasts or satellite models. This paper proposes a set of quantitative
metrics that can be used to summarize the overall discrepancy between two forecasting models jointly across variables and forecasting horizons. The methodologies can handle situations where only the point forecast is available as well as where the full predictive densities are known. It also allows to take into account the policymaker loss function, by assigning different weights to variables or horizons. We illustrate the usefulness of our measures when comparing the forecasts from the Survey of Professional Forecasters, the Tealbook, a medium scale Bayesian VAR, and a medium scale Dynamic Stochastic General Equilibrium (DSGE) model for the U.S. data. We find that the forecasts substantially depart ahead of and during recessions, resulting in our measures spiking.