POLIBITS, INSTITUTO POLITECNICO NACIONAL// (VOL. 55)
Volumen: 55, Numero: 1, Páginas: 49-57 pp.
Numerous statistical and mathematical methods have been developed in order to explain the complexity of nonstationary time series. Singular Spectrum Analysis (SSA) and Wavelet Transform (WT) are two potent theories with different mathematical foundations that have been used in several applications with successful results; however in most studies SSA and WT have been presented separately, then there is a lack of systematic comparisons between SSA and WT in time series forecasting. Consequently the aim of this work is to evaluate the performance of two hybrid models, one is based on SSA combined with the Autoregressive model (SSA-AR), and the other is based on Stationary Wavelet Transform combined with AR (SWT-AR). The models are described in two stages, the first stage is the time series preprocessing and the second is the prediction. In the preprocessing the low frequency component is obtained, and by difference the high frequency component is computed. Whereas in the prediction stage the components are used as input of the Autoregressive model. The empirical data applied in this study corresponds to the traffic accidents domain, they were daily collected in the Chilean metropolitan region from 2000 to 2014 and are classified by relevant causes; the data analysis reveals important information for road management and a challenge for forecasters by the nonstationary characteristics. The direct strategy was implemented for 7-days-ahead prediction, high accuracy was observed in the application of both models, SWT-AR reaches the best mean accuracy, while SSA-AR reaches the highest accuracy for farthest horizons.