EMD-based Hybrid Modeling Techniques For Time Series Forecasting And Their Applications

Empirical mode decomposition (EMD) is an ideal tool for time-frequency analysis of nonstationary signals. In recent years, it has been used successfully in many areas in science and engineering. However, its applications in nonlinear time series modeling and forecasting are still in infant stage. Despite that a limited studies have been reported in this area, the advantages of it are not fully exploited. In this regard, the goal of this dissertation is to go further with novel models and convincing empirical evidence to demonstrate the advantages of the EMD in nonlinear time series modeling and forecasting, particularly in fields of business and finance.The main contributions of this dissertation are as follows:First of all, this study proposes a hybrid modeling framework incorporating EEMD (Ensemble Empirical Mode Decomposition, a variant of EMD)and support vector regression for time series forecasting and conducts empirical study of stock price forecasting.Secondly, regarding to the end effect occurred during the sifting process of EMD, the research on EMD-based hybrid modeling framework has paid little, even no attention to. This study proposes improved EMD-based modeling framework incorporating selected end condition methods and then goes a step forward by comparing four mainstream end condition methods from the point of view of the prediction performance. Based on a large scale experimental datasets and comparison with counterparts, the experimental results demonstrated that the negative effect of end effect on the prediction performance of EMD-based hybrid modeling framework could be alleviated. Based on the above research, this study proposes an ensemble empirical mode decomposition (an updated version of EMD) based support vector machines (SVMs) modeling framework incorporating slope-based method to restrain the end effect issue to forecast air passenger traffic with high degree of volatility.Thirdly, Studies on EMD-based time series forecasting and prediction strategy.(1) This study proposes an improved PSO-MISMO modeling strategy for multi-step-ahead prediction that incorporates a heuristic based on particle swarm optimization (PSO) into the MISMO (multi-input several multi-output) modeling process to self-adaptively determine the number of sub-models with varying prediction horizons. For the purpose of justification, this study compares the rank of the proposed modeling strategy with the four well-established strategies with neural networks (NNs). Quantitative and comprehensive assessments are performed with the simulated and real time series on the basis of the prediction accuracy, convergence, and computational time.(2) Past studies on EMD-based modeling framework are limited to their preoccupation with one-step-ahead forecasting. This study proposes improved EMD-based modeling framework that applies to multi-step-ahead time series forecasting and conducts empirical study in the international crude oil price forecasting.(3) The traditional formulation of support vector regression (SVR) could be taken for multi-step-ahead time series prediction, only relying either on iterated strategy or direct strategy due to the single-output structure. This study proposes a novel multiple-step-ahead time series prediction approach which employs multiple-output support vector regression (M-SVR) with MIMO prediction strategy.Finally, past studies on time series forecasting are limited to their preoccupation with single-valued forecasting, this study proposes two interval-valued time series forecasting methods from two different research perspectives.(1) In the case of retaining the interval-valued data characteristic, by virtue of the multi-output structure of multiple-output support vector regression (MSVR) and high-efficient optimization ability of firefly algorithm (FA), this study proposes a FA-based MSVR hybrid model for interval-valued time series forecasting and conducts empirical study by combining with the interval-valued stock price index forecasting.(2) In the case of abandoning the interval-valued data characteristic, by virtue of the high-efficient decomposition ability for complex-valued series of bivariate empirical mode decomposition (BEMD) and the excellent performance of EMD-based modeling framework in single-valued time series forecasting, this study proposes BEMD-based modeling framework for interval-valued time series forecasting and conducts empirical study by combining with the interval-valued load demand forecasting.