|
Abstract:
|
Current methods for model-based seasonal adjustment tend to focus on selecting the optimal model (according to some criteria) and then computing the seasonal adjustment based on that unique model. Here, an alternative approach is proposed where the adjustment is based on weighted averages of the seasonally adjusted series from many models under consideration. Two main types of averaging are considered: frequentist (with likelihood-based weights) and Bayesian (with weights based on the Bayes factors of the models). The resulting averaging is compared in terms of its performance compared with current methods, both in terms of forecasting, and in terms of the revisions and seasonal stability of the adjusted series. This is investigated through both simulated series (where the true model may or may not be among the set considered) and real datasets.
|
