Aggregate versus disaggregate information in dynamic factor models

We examine the finite-sample performances of dynamic factor models that use either aggregate or disaggregate data, where the latter rely on finer disaggregations of the headline concepts of a small set of economic categories. Our Monte Carlo analysis reveals that the use of the series with the largest averaged within-category correlations outperforms the use of disaggregate data for factor estimation and forecasting in several cases. This occurs for high levels of cross-correlation across the idiosyncratic errors of series that belong to the same category, for oversampled categories, and especially for high levels of persistence in either the common factor or the idiosyncratic errors. However, the forecasting gains are reduced considerably when the target series are persistent. This could potentially explain why there is no clear ranking between the aggregate and disaggregate approaches when using the constituent balanced panel of the Stock-Watson factor model, which classifies the US data into 13 economic categories.