Traditional versus unobserved components methods to forecast quarterly national account aggregates

We aim to assess the ability of two alternative forecasting procedures to predict quarterly national account (QNA) aggregates. The application of Box–Jenkins techniques to observed data constitutes the basis of traditional ARIMA and transfer function methods (BJ methods). The alternative procedure exploits the information of unobserved high- and low-frequency components of time series (UC methods). An informal examination of empirical evidence suggests that the relationships between QNA aggregates and coincident indicators are often clearly different for diverse frequencies. Under these circumstances, a Monte Carlo experiment shows that UC methods significantly improve the forecasting accuracy of BJ procedures if coincident indicators play an important role in such predictions. Otherwise (i.e., under univariate procedures), BJ methods tend to be more accurate than the UC alternative, although the differences are small. We illustrate these findings with several applications from the Spanish economy with regard to industrial production, private consumption, business investment and exports.