Evaluating Pairwise Variable Selection Methods

Abstract

This paper discusses novel methods for the pairwise selection of explanatory variables from a large set of candidate pairs. These methods are applied to monthly time series of surface temperature and their performance is compared with that of conventional selection criteria such as AIC and BIC. In our frequency-domain analysis of the temperature datasets, the pairs are defined in a natural way as cosine and sine vectors of the same frequency. The results show that the new criteria are the only ones which are able to correctly identify seasonal patterns.