A mutual association based nonlinear ensemble mechanism for time series forecasting

Forecasting a time series with reasonable accuracy is an important but quite difficult task that has been attracting lots of research attention for many years. A widely approved fact is that combining forecasts from multiple models significantly improves the forecasting precision as well as often produces better forecasts than each constituent model. The existing literature is accumulated with linear methods of combining forecasts but nonlinear approaches have received very limited research attention, so far. This paper proposes a novel nonlinear forecasts combination mechanism in which the combined model is constructed from the individual forecasts and the mutual dependencies between pairs of forecasts. The individual forecasts are performed through three well recognized models, whereas five correlation measures are investigated for estimating the mutual association between two different forecasts.Empirical analysis with six real-world time series demonstrates that the proposed ensemble substantially reduces the forecasting errors and also outperforms each component model as well as other conventional linear combination methods, in terms of out-of-sample forecasting accuracy.