First, because empirical evidence suggests that time variation in macroeconomic relations is widespread (e.g. Stock and Watson [1996]), the factor loadings are permitted to evolve over time. Second, the factor structure is approximate, in the sense that the idiosyncratic errors can be correlated across series. Third, the model is nonparametric, in the sense that the correlation structures and distributions of the idiosyncratic terms and the factors, and the precise lag structure by which the factors enter, are not specified parametrically. Fourth, a practical concern when working with a large number of time series is that a large break or outlier arising from a data entry error or a redefinition might go undetected, and this possibility is introduced into the analysis. Fifth, because economic time series are typically available over different spans, the model and estimation procedures are developed for the cases of both a balanced and unbalanced panel.

This factor-based approach to forecasting can be contrasted to conventional regression-based model selection. With a very large number of predictor variables, it is computationally infeasible to enumerate and to estimate all possible models up to a given order. Although this computational problem can be ameliorated by making informed choices about the models to be estimated, more fundamentally the model selection approach is prone to producing particularly poor out of sample forecasts because of fitting so many models. In contrast, the dynamic factor model places very strong restrictions on the joint behavior of the predictors that permits extreme parameter reduction, so that for forecasting purposes the very many predictors can be replaced by a handful of factors.