Description

This paper evaluates the performance of a suite of traditional models used to forecast short-term quarterly GDP, going from SARIMA, BVAR, dynamic factors, Bridge models to MIDAS. In total, 155 specifications are considered, and the accuracy of the forecasts is evaluated by means of a rolling out-of-sample prediction exercise for a four-quarter horizon. The main results suggest that forecasts improve as information from the current quarter is incorporated. Also that IMACEC is particularly useful given that it allows expressing GDP in monthly frequency. And finally, that relative performance of models can change abruptly with economic conditions, meaning that combinations of models tend to outperform most of the individual models, which is consistent with the literature.