Local Learning for Iterated Time-Series Prediction

Abstract

We introduce and discuss a local method to learn one-step-ahead predictors for iterated time series forecasting. For each single one-stepahead prediction, our method selects among different alternatives a local model representation on the basis of a local cross-validation procedure. In the literature, local learning is generally used for function estimation tasks which do not take temporal behaviors into account. Our technique extends this approach to the problem of long-horizon prediction by proposing a local model selection based on an iterated version of the PRESS leave-one-out statistic. In order to show the effectiveness of our method, we present the results obtained on two time series from the Santa Fe competition and on a time series proposed in a recent international contest. 1 Introduction The use of local memory-based approximators for time series analysis has been the focus of numerous studies in the literature [5, 14]. Memory-based approaches do not estimate a global model...