Boot.EXPOS in NNGC competition

In this article the authors expose an automatic procedure that combines a very popular resampling technique, the Bootstrap methodology, with one of the most widely used forecasting methods, the exponential smoothing. The merge of these two approaches originates the Boot.EXPOS. The algorithm can be summarized as follow: Given a time series, it starts by selecting the “best” EXPOS model for fitting the data, using the AIC criterion. The fitted values and the estimated EXPOS parameters are kept for later reconstructing the time series. Concerning the random part, a parametric model is fitted and then the residuals are bootstrapped. Here Boot.EXPOS procedure is applied. It can be summarized as follows: (i) fit an AR(p), where the order p is selected by AIC criterion; (ii) obtain the AR residuals; center the residuals; draw a sample with replacement from the centered residuals; (iii) obtain a bootstrapped replica recursively from the AR adjustment and using the bootstrap residuals series from the previous step; (iv) join the bootstrapped replica and the fitted EXPOS model to obtain a sample path of the initial time series; (v) use the new time series to obtain the EXPOS forecasts and forecast intervals. An additional algorithm NABoot.EXPOS was developed to handle with missing data. It detects, estimates and replaces the unobservable values.