Abstract

In research of time series forecasting, a lot of uncertainty is still related to the question of which forecasting method to use in which situation. One thing is obvious: There is no single method that performs best on all time series. This work examines whether features extracted from time series can be exploited for a better understanding of different behaviour of forecasting algorithms. An extensive pool of automatically computable features is identified, which is submitted to feature selection algorithms. Finally, a possible relationship between these features and the performance of forecasting and forecast combination methods for the particular series is investigated.

Document Downloads