Generalized Exponential Weighted Moving Average for Time Series Forecasting

• Activity Number: 424
• Type: Contributed
• Date/Time: Tuesday, August 3, 2010 : 2:00 PM to 3:50 PM
• Sponsor: Section on Nonparametric Statistics
• Author(s): Lu Wang*+
• Companies: University of California, Davis
• Keywords: time series prediction ; exponential smoothing ; autoregressive integrated moving average ; state-space model

Abstract

We aim to provide a new methodology for predicting univariate time series and time series regression. This method is motivated by the ordinary exponential weighted moving average (EWMA). A general class of predictors based on a generalization that involving taking arbitrary linear combinations of EWMA estimators is constructed and is shown to be dense in the class of infinite order autoregressive process. Practical implementation of the method is simple after we recast the model in the usual Gaussian-Markov framework. We applied several strategies for model fitting and forecasting, e.g. stepwise regression, ridge regression, penalized least squares. Simulation and data analysis indicate that our proposed procedure may outperform the existing methods. Similar methodology has also been applied to time series regression with independent predictors.