Title: Combining long-run survey forecasts and nowcasts with VAR forecasts using relative entropy
Authors: Saeed Zaman - Federal Reserve Bank of Cleveland (United States)
Ellis Tallman - Federal Reserve Bank of Cleveland (United States)
Ellis Tallman - Federal Reserve Bank of Cleveland (United States) [presenting]
Abstract: Previous research highlights how nowcasts can improve the forecast accuracy in both Vector Autoregressions (VARs) and Dynamic Stochastic General Equilibrium (DSGE) models. Research has also highlighted the superior performance of the long-horizon survey forecasts compared to econometric approaches to forecasting long-horizon trajectories. We use real-time forecast evaluation to show that combining both long and short term conditions within one modeling approach generates meaningful gains in forecast accuracy. Specifically, we combine VAR forecasts with both external nowcasts and long-horizon survey forecasts using relative entropy to refine the medium-term forecasts of the VAR. The horizon at which we combine the VAR forecast with the long-horizon survey forecast varies by variable depending upon the degree of persistence of the variable. We propose a simple method to determine the relevant horizon for combination. Given that surveys are performed infrequently and do not cover all the forecast horizons, our procedure can also be thought as an approach to interpolate survey forecasts.