Abstract
This thesis makes three distinct contributions to the literature on factor-augmented models for forecasting economic time series using big datasets. The first chapter extends Diebold-Mariano-West type tests of forecast accuracy to apply to factor-augmented models where both factors and model coefficients are estimated in a rolling out-of-sample estimation procedure. This set-up poses new challenges as the sign of neither the factors nor factor-augmented model parameters are identified in different rolling windows. We propose a novel new identification strategy which removes arbitrary sign-changing in the sequence of out-of-sample parameter estimates and allows us to establish the asymptotic normality of the Diebold-Mariano test statistic. We propose a new bootstrap procedure for rolling factor estimates as existing bootstrap methods cannot deal with the generated regressor structure of the factors. The second chapter provides consistent information criteria for the selection of forecasting models which use both the idiosyncratic and common factor components of a big dataset. This procedure differs to existing factor-augmented model selection techniques as it depends on estimates of both the factors and the idiosyncratic components. We show that the combined estimation error vanishes at a slower rate than in the case of pure factor-augmented models in most standard economic forecasting scenarios, which makes existing information criteria inconsistent. We solve this problem by proposing modified information criteria which account for the additional source of estimation error. The final chapter aims to improve factor-based forecasts by 'targeting' factor estimates with two objectives: (i) so they are more relevant for a specific target variable, and (ii) so that variables with high levels of idiosyncratic noise are down-weighted prior to factor estimation. Existing targeted factor methodologies are only capable of estimating factors with one of these two objectives in mind. We suggest new Weighted Principal Components Analysis (WPCA) and Targeted Generalized PCA (TGPCA) procedures, which both use LASSO-type pre-selection.