A Novel Parallel Hybrid Model Based on Series Hybrid Models of ARIMA and ANN Models

摘要

The complexity of real-world time series makes to hardly yield the desired prediction performance by the existing individual models. A series hybrid model that relies on decomposing time series and then sequentially modeling patterns have been frequently adopted in this domain. Nevertheless, a lack of capturing all existing patterns, including all mixed and pure ones, may lead to a decrease in the performance of series hybrid models. Thus, this study aims to propose a novel parallel hybrid model to achieve a comprehensive hybrid framework in which all pure and mixed linear and/or nonlinear patterns in real-world time series can be appropriately modeled. Based on the data generation process, four modeling procedures, including pure linear and nonlinear, linear-nonlinear, and nonlinear-linear modeling structures, can be considered in data. Therefore, this study proposed a hybrid system contains five main parts (1) linear modeling using ARIMA model, (2) nonlinear modeling using MLPNN model, (3) linear-nonlinear sequential modeling using series ARIMA-MLPNN hybrid model, (4) nonlinear-linear sequential modeling using series MLPNN- ARIMA hybrid model and (5) integrating the four obtained forecasts employing a parallel hybridization. Three real-world data sets from the international stock market namely closing of the DAX index, the closing of the Nikkei 225 index (N225), and the opening of the Dow Jones Industrial Average Index are utilized to demonstrate the performance of the proposed hybrid model. The proposed hybrid model outperforms ARIMA, MLPNN, RBFNN, LSTM individual models as well as ARIMA-MLPNN, MLPNN-ARIMA series hybrid models, and parallel hybridization of ARIMA and MLP models.