Forecasting using a large number of predictors: Is Bayesian shrinkage a valid alternative to principal components?