This thesis proposes a method that uses dynamic selection to construct ensembles for time series forecasting. Dynamic selection is used to construct the ensemble for each prediction depending on the given example. The idea behind using the dynamic selection is to make the ensemble adaptable to the changes of the time series and the changes of each models performances. Besides selecting an ensemble, dynamic selection can be used to select a single model for each prediction. Selecting only a single model would neglect the improvements that ensemble methods can deliver in terms of making the most of multiple models.