In this paper the author presents the theoretical background for ensemble methods via multivariate decompositions. Having a set of models he treats their results as a multivariate variable with destructive and constructive latent components. The data decorrelation and MSE error decompositions provide to proper destructive components identification and elimination. The elimination of destructive components should improve final prediction. The validity of the approach is verified by energy load prediction problem.
Ryszard Szupiluk, Szkola Glówna Handlowa w Warszawie, Kolegium Analiz Ekonomicznych, al. Niepodleglosci 162, 02-554 Warszawa, Poland
Publication order reference
CEJSH db identifier