CMStatistics 2018: Start Registration
View Submission - CMStatistics
B1668
Title: A contribution to forecast time series with structural break Authors:  M Rosario Ramos - FCiencias.ID (Portugal) [presenting]
Clara Cordeiro - CEAUL and FCT, UALg (Portugal)
Abstract: The presence of structural instability affects the estimation, inference and prediction. In this case, for a time series $Y= {y_1,\cdots, y_n}$, a structural change exists at a unknown time t if $y_1, \cdots , y_t$ differ from $Y^*={y_{t+1},\cdots, y_n}$, namely in the trend. This has an impact on the study of economic, environmental, climatic and other variables, since the forecast can be strongly biased. The aim is to contribute for the improvement of forecasts in this scenario. Given a time series $Y$, the first step is the estimation and removing of the seasonality based on the seasonal-trend decomposition by Loess (STL). The selection of the best STL fit was performed by the algorithm stl.fit, which runs all the possible combinations of the smoothing parameters and in the end, find the optimal parameters combination which minimized an accuracy measure. Secondly, the detention of a structural break in the seasonally adjusted time series is performed by the R package strucchange. Thirdly, based on the time series $Y^*$, fit the best model and obtain forecasts. A comparison between the forecasts of $Y$ and $Y^*$ is presented and the performance is evaluated using real data sets.