Predictive performance of some nonparametric linear and nonl-inear smoothers for noisy data

Abstract

The purpose of this study is to discuss the weighting systems of several linear and non linear smoothers and to evaluate their predictive performances when applied to noisy time series. On this regard, we illustrate with three Canadian leading indicators which are representative of larger sets of time series characterised by a low, medium and high signal to noise ratio. The smoothers discussed are: (a) loess (a locally weighted regression smoother), (b) Gaussian Kernel smoother, (c) supersmoother, (d) cubic smoothing spline and (e) Dagum’s modified 13-term Henderson filter. Their performances are evaluated on the basis of three following main criteria for current economic analysis: (1) number of unwanted ripples or false turning points in the final estimated trend, (2) time lag in detecting ‘true’ turning points and (3) size of total revision of the concurrent trend estimates.