Automatic Model Selection in Ensembles for Time Series Forecasting

Long-term forecasting in time series is still an open problem, but promising advances have been achieved in this area. Among them, it has been found that the best predictions may be obtained when combining different forecasting models. In this context, diversity and accuracy of the involved models are the most important factors to be considered when selecting them. In this paper, we analyze the results of a new method for multiple-step prediction, based on a Self-Organizing Map (SOM) neural network and meta-features. Using a rule of pruning, this method automatically adjusts the required balance between diversity and accuracy in the selection of the forecasters. The method was tested for the prediction of long term horizons, using synthetic and real time series produced by highly nonlinear systems. Our results showed that, on average, this method obtains better forecasting results than the results obtained using other state-of-the-art methods.