Model selection and randomization for weakly dependent time series forecasting

Observing a stationary time series, we propose in this paper new two steps procedures for predicting the next value of the time series. Following machine learning theory paradigm, the first step consists in determining randomized estimators, or "experts", in (possibly numerous) different predictive models. In the second step estimators are obtained by model selection or randomization associated with exponential weights of these experts. We prove Oracle inequalities for both estimators and provide some applications for linear, artificial Neural Networks and additive non-parametric predictors.ou