Time Series Models And Forecasting Based On Computational Intelligence

Time series forecasting is a hot research topic. Autoregressive integrated moving average is one of the most classical time series models which has been proposed by Box G. E. P. and Jenkins G. M. in the early 70s. But it has two limitations, one is the assumption of linear structure, and the other is the requirement for more observations in order to obtain ideal forecasting. In practice, there may be nonlinear structure characteristics and missing data in time series. In addition, time series models are often established using the information itself, but there are few studies focusing on the integrating of relevant information. In this thesis, we propose some time series forecasting models based on computational intelligence technology and Granger causality. The main research works are as follows:1. A fuzzy time series forecasting model using fuzzy clustering and information granule is proposed. Partitioning the universe of discourse and determining effective intervals are critical for forecasting in fuzzy time series. Equal length partition is subjective in practice, even if it is used in most existing literatures because of its convenience. At this point, we study how to partition the universe of discourse into intervals with unequal length. Firstly, we calculate the prototypes of data using fuzzy clustering algorithm. Secondly, form some subsets according to the prototypes. Finally, the unequal length intervals are obtained by information granule. The method of partitioning fully considers the distribution of time series observation and these intervals carry well-defined semantics. To verify the suitability and effectiveness of the approach, we apply the proposed model to forecast enrollment of students of Alabama University and Germany's DAX stock index monthly values. Empirical results show that the proposed model with unequal length intervals can greatly improve forecast accuracy.2. This thesis presents a fuzzy time series forecasting model containing useful temporal information. Temporal information plays an important role for time series prediction. In the proposed model, time variable is involved in partitioning the universe through Gath-Geva clustering algorithm. The authors not only consider the data distribution but also take the effect of time variable into account to the partitioning of the universe of discourse. As a result, the proposed model contains more information and can well reflect the changing trends of time series. We apply the proposed method to forecast enrollment of students of Alabama University and the Taiwan stock exchange capitalization weighted stock index. The experimental results verify the superiority of the proposed model.3. This study presents a different approach to time series forecasting using a concept of Granger causality. Granger causality relationships between variables of interest and relevant variables are involved in the formation of time series forecasting model. First, we construct forecasting model only using information about variables of interest themselves. Second, we utilize hypothesis testing to determine the Granger causality relationships between variable of interest and relevant variables and capture the functional relationship between variables exhibiting Granger causality. Finally, the Granger causality relationships and the information of time series itself are integrated into the final time series forecasting model. The proposed model fully considers the dynamic characters among time series variables. We demonstrate the quality of the proposed method using some synthetic data and a real world three-dimensional time series. The obtained experimental results quantify the effectiveness of the proposed approach.