| Abstrakt: |
Parallel hybridization is one of the most well-established hybrid structures proposed in the literature. Since an unavoidable high degree of multi-collinearity (MC) exists among predictors in hybrid structures, the hybrid results and estimated parameters may not be generalizable and stable. Thus, the major innovation of this study lies in addressing this shortcoming that parallel hybrid models faced and eliminating the MC problem, which is not yet addressed in the literature on time series forecasting using hybridization methodologies. To solve this difficulty and analyze the effect of the MC phenomenon on the forecasting accuracy of the parallel hybrid model as well as the reliance degree of estimated weights and results, a new class of hybrid models is proposed based on parallel integration of principal component analysis (PCA), auto-regressive integrated moving average (ARIMA), multi-layer perceptron neural network (MLPNN) and exponential smoothing model (ESM). The forecasting accuracy and reliance degree of the proposed model is compared with the traditional parallel hybridization of ARIMA, MLPNN, and ESM models. The experimental results revealed that by removing MC, improved forecasting accuracy is obtained. Besides, the reliability and statistical power of results, specifically the estimated weights, are enhanced. The verification results indicate that the proposed PCA-ARIMA&MLP&ESM model is a powerful applicable tool for time series forecasting in terms of forecasting accuracy and model reliability. |
Full text from SpringerLink