Forecasting macroeconomic variables using large datasets : dynamic factor model versus large-scale BVARs

Abstract

This paper uses two-types of large-scale models, namely the Dynamic Factor Model (DFM) and Bayesian Vector Autoregressive (BVAR) Models based on alternative hyperparameters specifying the prior, which accommodates 267 macroeconomic time series, to forecast key macroeconomic variables of a small open economy. Using South Africa as a case study and per capita growth rate, inflation rate, and the short-term nominal interest rate as our variables of interest, we estimate the two-types of models over the period 1980Q1 to 2006Q4, and forecast one- to four-quarters-ahead over the 24-quarters out-of-sample horizon of 2001Q1 to 2006Q4. The forecast performances of the two large-scale models are compared with each other, and also with an unrestricted three-variable Vector Autoregressive (VAR) and BVAR models, with identical hyperparameter values as the large-scale BVARs. The results, based on the average Root Mean Squared Errors (RMSEs), indicate that the large-scale models are better-suited for forecasting the three macroeconomic variables of our choice, and amongst the two types of large-scale models, the DFM holds the edge.