A1886
Title: Forecasting with exponential smoothing models using bootstrapped model selection and parameter estimation
Authors: Livio Fenga - University of Exeter (United Kingdom) [presenting]
Abstract: Widely used in the field of time series analysis for a variety of tasks (e.g. forecasting and simulation), exponential smoothing models are recognized as a powerful tool adopted in many contexts (applied research, official bureaus, public and private companies) by many actors (e.g. statisticians, econometricians and practitioners). Regardless of the purpose exponential smoothing models are built for, their usefulness greatly depends on the goodness of their parameters' estimates. The related inference procedures are, in many instances, carried out under the maximum likelihood paradigm, which, unfortunately, can be heavily impacted by different sources of errors, induced by bias components and uncertainty. The present paper outlines a computer-intensive procedure, aimed at attenuating the effects of such errors. The proposed approach, based on a bootstrap scheme of the type maximum entropy, is theoretically discussed and empirically evaluated in terms of forecasting performances within the minimum Akaike information criterion expectation (MAICE) framework, using the 366 monthly time series from the M3 2010 tourism forecasting competition dataset, freely and publicly available in the R package Tcomp.
