Comparison of Model Reduction Methods for VAR Processes

The objective of this study is to compare alternative computerized model-selection strategies in the context of the vector autoregressive (VAR) modeling framework. The focus is on a comparison of subset modeling strategies with the general-to-specific reduction approach automated by PcGets. Different measures of the possible gains of model selection are considered: (i) the chances of finding the correct model, that is, a model which contains all necessary right-hand side variables and is as parsimonious as possible, (ii) the accuracy of the implied impulse-responses and (iii) the forecast performance of the models obtained with different specification algorithms. In the Monte Carlo experiments, the procedures recover the DGP specification from a large VAR with anticipated size and power close to commencing from the DGP itself when evaluated at the empirical size. We find that subset strategies and PcGets are close competitors in many respects, with the forecast comparison indicating a clear advantage of the PcGets algorithm.