Study On Combination Forecasting Model For Time Series Based On SARIMA And BP Neural Network

Time series are common in some areas of research such as finance, transportation and engineering fields, in which many theoretical and practical issues have an urgent need for the analysis of time series. The actual time series affected by many different factors includes the important information reflecting the relationship between the sequence and the sequence of their cycle fluctuations so we need to seriously study the role of these factors.Since the1900s, scholars have been persistently making analysis of seasonal time series. They put forwards a variety of models, such as SARIMA model, X12-ARIMA model, BP neural network model, GM(1,1),and so on. These efforts make further development of the theory and application of the time series.After a comparative analysis of various models, we find that different prediction models have advantages and disadvantages and reflect some of the information of the original data. The SARIMA model, one of the classical models, is a generalization of the ARIMA model in seasonal time series and has a strong capability on linear modeling. BP neural network model, one of the artificial intelligence methods recently, is also used to solve the problelem of time series because of its good ability on nonlinear model prediction. In order to improve the prediction accuracy further and make full use of the advantages of different models, Bates and other researchers put forwards the idea of combined model. Firstly, we introduce the particle swarm optimization (PSO) algorithm to search the optimal weight coefficients of each individual prediction model and propose the combination forecasting model based on SARIMA and BP neural network. Secondly, with reference to adaboost combined classified algorithm, a model based on BP_Adaboost is proposed for time series prediction. This paper applied the combined model to the analysis of time series of China’s total retail sales of social consumer goods and the effects of different models are compared. The results showed that the combination forecasting model is superior to single prediction model and has better predictive ability...