Parsimonious time series modeling for high frequency climate data

Abstract

Climate data often provides a periodically stationary time series, due to seasonal variations in the mean and covariance structure. Periodic ARMA models, where the parameters vary with the season, capture the nonstationary behavior. High frequency data collected weekly or daily results in a large number of model parameters. In this article, we apply discrete Fourier transforms to the parameter vectors, and develop a test for the statistically significant harmonics. An example of daily high temperatures illustrates the method, whereby a periodic autoregressive model with 1095 parameters is reduced to a parsimonious 12 parameter version without any apparent loss of fidelity.