Abstract
Factor models (FM) are now widely used for forecasting with large set of time series. Another class of models, which can be easily estimated and used in a large dimensional setting, is multivariate autoregressive models (MAR), where independent autoregressive processes are assumed for the series in the panel. When applied to big data, the estimation, model selection and combination of both models can be time consuming. We assume both FM and MAR models are misspecified and provide a scoring rule which can be evaluated on an initial training sample to either select or combine the models in forecasting exercises on the whole sample. Some numerical illustrations are provided both on simulated data and on well known large economic datasets. The empirical results show that the frequency of the true positive signals is larger when FM and MAR forecasting performances differ substantially and it decreases as the horizon increases.