A1986
Title: Evaluating forecast performance at the Bank of England
Authors: Emi Mise - University of Leicester (United Kingdom) [presenting]
Anthony Garratt - University of Warwick (United Kingdom)
Abstract: The Bank of England has historically reported predictive densities for key macroeconomic variables in the form of split-normal distributions. This practice was criticised in the Bernanke report, which evaluated the Banks forecasting performance. A Bayesian time-varying stochastic VAR model is developed, building on the specification proposed by Clark, McCracken and Merten (CMM-SV). The approach uses nowcast forecast errors and exploits the accounting identity between prediction errors and revisions to forecasts at different horizons in order to model time-varying uncertainty. Real-time predictive densities for U.K. macroeconomic variables are constructed and evaluated against the Bank's split-normals, a simple normal error benchmark, and densities derived from the Survey of External Forecasters, using evaluation methods including coverage rates, continuous ranked probability score (CRPS) and probability events of interest. The analysis also jointly models forecast errors and forecast revisions for inflation and output growth. Results indicate that this method, which delivers sizeable gains for U.S. data relative to prediction intervals based on historical forecast-error variances, including during periods of heightened uncertainty, yields only limited gains over the Bank's split-normal distributions and other institutional benchmarks. These findings suggest that the benefits of introducing stochastic volatility for density prediction may be context-dependent.
