Title: A Bayesian framework for handling outliers in seasonal adjustment
Authors: Anindya Roy - University of Maryland Baltimore County (United States) [presenting]
Tucker McElroy - Census Bureau (United States)
Abstract: Current seasonal adjustment approaches require the identification of extreme values and outliers as fixed effects as an initial step, followed by their removal. The extreme value adjusted series are then filtered using linear techniques ideally suited for Gaussian observations. The final analysis however ignores the added uncertainty due to the specification of the time of occurrences of the outliers and also the effect of inference from the removal of the outliers. Alternatively, the outliers can be modeled as arising from latent stochastic processes driven by heavy-tailed innovations; extraction of latent components then follows non-linear techniques, and does not require identification of extreme epochs. We propose a Bayesian framework for modeling the dynamic components of a time series along with the extreme observations. The approach also yields a posterior for a counterfactual trend estimate that helps evaluate the impact of extreme occurrences.