Forecasting with nostationary dynamic factor models

Abstract

In this paper we analyze the structure and the forecasting performance of the dynamic factor model. It is shown that the forecasts obtained by the factor model imply shrinkage pooling terms, similar to the ones obtained from hierarchical Bayesian models that have been applied successfully in the econometric literature. Thus, the results obtained in this paper provide an additional justification for these and other types of pooling procedures. The expected decrease in MSE f or using a factor model versus univariate ARIMA models, shrinkage univariate models or vector ARMA models are studied f or the one factor model. It is proved that some substantial gains can be obtained in some cases with respect to the univariate forecasting. Monte Carlo simulations are presented to illustrate this result. A factor model is built to forecast GNP of European countries and it is shown that the factor model provides better forecasts than both univariate and shrinkage univariate forecasts.