The generalization ability is one of the most important and the most influential factors for electing forecasting models, managing future events, and making decisions. In the literature, numerous hybrid models have been presented in order to improve the accuracy as well as generalization ability of single forecasting approaches. The main aim of these hybrid models is often to use more different and/or more individual models in order to capture all existing patterns and structures in the data, more completely; and consequently improving the accuracy and generalization. Although, it can be generally demonstrated that increasing the number of components will not decrease the performance of hybrid models in the training, it will not necessarily improve the generalizability, especially in complex and uncertain environments. In this paper, an efficient allocation strategy is proposed in order to assign the underlying data set to its appropriateness component for increasing generalizability as well as decreasing computational costs. In this paper, a novel soft intelligent hybrid model is developed using the allocation strategy for assign different IMFs to appropriateness certain linear, certain nonlinear, uncertain linear, and uncertain nonlinear components in decomposition based forecasting problems. The main purpose of this classification is to reduce the probability of the over-fitting problem and consequently to increase the generalization ability, in additional of deceasing the computational costs. Moreover, in this paper, an optimal weighting technique is proposed to find the relative importance of each component in order to yield the most accurate final predictions. On the other hand, the main motivation of the paper, in contrast to the regular decomposition based hybrid models in which components are blindly assigned to the models, is to develop a logical process to allocate components to the most appropriate model as well as optimally weighting them. Empirical results of crude oil prices and wind power forecasting indicate that despite of better performance of traditional parallel hybrid models in the training sample, the generalization ability of the proposed model in test sample is significantly higher than those hybrid models as well as its components in all considered benchmarks. The proposed model can averagely improve 64.86%, 61.93%, and 52.00% the accuracy of single linear, single nonlinear, and traditional hybrid non-decomposition; and 41.37%, 35.16%, and 32.63% the performance of single linear, single nonlinear, and traditional hybrid decomposition based models, respectively.
Read full abstract