Title: Automatic algorithms for univariate time series forecasting using SARIMA and hybrid Wavelet-ARIMA-neural networks models
Authors: Dedi Rosadi - Universitas Gadjah Mada (Indonesia) [presenting]
Abstract: In some application of time series modeling, it is necessary to obtain forecast of various types of data automatically and possibly in real-time, for instance, to do a (near) real-time processing of the satellite data. Various automatic algorithms for modeling univariate time series data are available in the literature. One class of the methods is the automatic algorithm to model and to forecast univariate SARIMA models. We discuss three methods of these types of algorithms, one of them based on a combination between the best exponential smoothing models to obtain the forecast, together with state-space approach of the underlying model to obtain the prediction interval. Other method, which is more advanced method, is based on X-13-ARIMA-SEATS. Other method use more heuristic approaches, namely the genetic algorithms. We also consider other general approaches and they are not restricted only to model the class of SARIMA models. One of the methods is based on automatic neural network method. The other (new) automatic algorithms are hybrid methods, which we called automatic Zhang (which uses a decomposition of the data into linear and nonlinear part), automatic KAV method (which combine decomposition of the data using wavelet and Zhang approach) and other type of automatic KAV method (which uses different architecture of the model). All of these approaches are implemented in our R-GUI package, namely RcmdrPlugin.SPSS. We provide empirical application using real data.