Title: Use of panel VAR models for nowcasting GDP
Authors: Dan Rieser - European Commission, Eurostat (Luxembourg) [presenting]
Gian Luigi Mazzi - Independent Expert (Luxembourg)
James Mitchell - University of Warwick (United Kingdom)
Abstract: The application of Panel Vector Auto-Regressive (PVAR) models to macroeconomic nowcasting is examined. Alternative model specifications proposed in the literature such as large scale Bayesian VAR models, Global VAR models and Dynamic Factor Models are reviewed. Their characteristics and performance for both forecasting and nowcasting is assessed by nowcasting GDP growth for 8 European countries and the Euro area. A mixed frequency, cross-country dataset is implemented for this purpose. This is a novelty as the use of mixed-frequency data (quarterly and monthly) in a PVAR model has previously been untried. In the past, nowcasts have been produced ignoring cross-country dependencies. The results based on six alternative panel VAR models point and density estimates suggest that the use of panel VAR model, with panel priors, can be helpful when forecasting GDP growth. However, when nowcasting based on the panel structure of the dataset and estimating a panel VAR model with panel priors, panel VAR models do not seem to be the preferred option. The most accurate nowcasts are obtained by shrinking the VAR model to a univariate model. This can be achieved by deploying the Minnesota prior that ignores the panel structure of the dataset and by not motivating the prior with reference to possible dynamic interdependencies, static interdependencies and cross sectional heterogeneities that one might expect in a panel dataset.