Comparison of Bayesian Moving Average and Principal Component Forecasts for Large Dimensional Factor Models

The growing availability of financial and macroeconomic data sets including a large number of time series (hence the high dimensionality) calls for econometric methods providing a convenient and parsimonious representation of the covariance structure both in the time and the cross-sectional dimensions. Currently, dynamic factor models constitute the dominant framework across many disciplines for formal compression of information. To overcome the challenges of dimensionality, many forecast approaches proceed by somehow reducing the number of predictors. Principal component regression (PCR) approach proposes computing forecasts as projection on the first few principal components of the predictors. Bayesian model averaging (BMA) approach combines forecasts to extract information from different possible relationships between the predicted variable and the predictor variables. These two literature apparently moved in two different directions. However, recent findings by De Mol et al. [2008] and the Ouysse and Kohn [2009] suggest there are theoretical and practical reasons to connect the two literatures. This paper provides empirical evidence for connecting these two seemingly different approaches to forecasting. The empirical results serve as a preliminary guide to understanding the behaviour of BMA under double asymptotics, i.e. when the cross-section and the sample size become large.