Title: A bootstrap contribution in STL decomposition
Authors: Clara Cordeiro - CEAUL and FCT, UALg (Portugal) [presenting]
Manuela Neves - FCiencias.ID, Universidade de Lisboa and CEAUL (Portugal)
M Rosario Ramos - FCiencias.ID (Portugal)
Abstract: A time series is the result of observing the values of a variable over time during regular intervals and can be seen as the result of the combination of different components. The classical methods of decomposition of time series allow to identify the trend, the seasonality and the irregular components. However, these methodologies do not allow for a flexible specification of the seasonal component, and the trend component is generally represented by a deterministic time function, which is easily affected by the existence of outliers. The nonparametric Seasonal-Trend decomposition by Loess (STL) is able to identify a seasonal component that changes over time, a non-linear trend and it can be robust in the presence of outliers. Bootstrap methods, initially introduced for independent random variables, can be successfully applied to time series. The Boot.EXPOS procedure, joining bootstrap and exponential smoothing methods, has revealed promising results. The aim is to explore the use of the Boot.EXPOS in predicting the components of the STL decomposition. Based on an error measure, the best STL fit is chosen. In the case of an uncorrelated irregular component, the forecast of the STL will rely on the forecast of trend and seasonal components obtained through the Boot.EXPOS.