The study of a time series has forecasting as one of its primary objectives. Exponential smoothing methods (EXPOS) stand out due to their versatility in the wide choice of models that they include. The widespread dissemination makes them the most widely used methods of modeling and forecasting in time series. An area that has given great support to the statistical inference is computational statistics, specifically the bootstrap methodology. In time series that methodology is most frequently used through the residual resampling. An automatic procedure that combines exponential smoothing methods and the bootstrap methodology was developed in environment. This procedure (Boot.EXPOS) selects the most appropriate model among a wide range of models, and performs an autoregressive (AR) adjustment to the EXPOS residuals. Once the stationarity of the residuals has been guaranteed, the AR residuals are resampled and the reconstruction of the original series is performed using the estimated components of the initial model. Point forecasts and prediction intervals are also provided. NABoot.EXPOS is an extension of that procedure that allows for the detection, estimation and imputation of missing values. An exhaustive study of several types of real time series given in competitions is presented in order to compare our procedures.

Doutoramento em Matemática e Estatística - Instituto Superior de Agronomia