Model Confidence Sets for Lag Selection in Autoregressive Models
We investigate the usefulness of the model confidence set algorithm proposed by Hansen, Lunde, and Nason (2011) for linear univariate time series modeling. We generate time series from simple AR(3) processes and conduct full model searches within the class of AR models of finite order, including subset models. Using squared error loss from out-of-sample predictions, we find that the power of the MCS procedure is very low when measured against the set of superior models which has exactly one element. This set ignores models that exhibit almost as low mean squared errors of prediction as the true model does. We therefore define a set of practically superior models which allows to include models with marginally worse performance where the performance difference is practically irrelevant. It turns out that in larger samples the MCS procedure identifies this set of practically superior models very well if the signal-to-noise ratio is sufficiently large. We then compare the performance of the MCS procedure to classical model selection criteria in terms of estimating impulse responses and obtaining forecasts for multiple horizons. Our results are ambiguous.
16.04.2015 | 17:00 c.t.
Kaminzimmer (Raum 202), Boltzmannstraße 20, Berlin-Dahlem