Tracking the best time series model

Mixture of experts algorithms have been shown to achieve a total loss close to the total loss of the best expert over a sequence of examples. We consider the use of the mixture of experts algorithms to track the best parameter settings of a predictor of a time series, especially when the time series is non-stationary. In this paper we introduce a new variant of the fixed-share algorithm "a mixture of experts algorithm" which we call the fixed-share hierarchy (FSH). We demonstrate the successful use of the FSH algorithm in tracking the best mixture of smoothing parameters of the Holt-Winters predictor when applied to a number of weather data sets. We show in a number of experiments that the FSH algorithm had lower total losses compared to the original fixed-share algorithm