ASYMPTOTIC PRINCIPAL COMPONENTS ESTIMATION OF LARGE FACTOR MODELS

There has been much recent interest in forecasting based on factor analysis models for large numbers of observable variables (p) and large numbers of observations (T). Some nice asymptotic results have been produced showing that under certain conditions, as (p,T) ? (8, 8) principal components analysis can be used to carry out the forecasting, thereby avoiding the need to fit a full factor analysis model. However, the question of how large p needs to be in order for the asymptotic theory to provide an adequate approximation in practice is open. In this paper we develop probability bounds for the forecast accuracy of principal component forecasts for stationary processes in terms of an empirically determinable noise to signal ratio. We develop a hypothesis test for this bound for which asymptotics in T hold even with p large and apply this test to US macrodata.