Component versus tradicional models to forecast quarterly national account aggregates: a Monte Carlo experiment

Abstract

Econometric models applied to observed data, specified and estimated using traditional Box-Jenkins techniques, have been widely used to forecast Quarterly National Account (QNA) aggregates. We assess the extent to which an alternative forecasting procedure, based on component models, improves the forecasting accuracy of traditional methods. Component models distinguish between the stochastic processes underlying the low- and the high-frequency component of time series, while traditional methods do not. Relationships between QNA aggregates and their coincident indicators are often significantly different for diverse frequencies, as suggested by even an informal examination of empirical evidence. Under these circumstances, a Monte Carlo out-of-sample experiment reveals that component models improve the forecasting accuracy of traditional methods to predict QNA aggregates when their coincident indicators play an important role in such predictions. Otherwise, specially when dealing with pure univariate specifications, traditional procedures likely beat component methods. We illustrate these findings with several applications for the Spanish economy.