Identification of time series components using break for time series components (bftsc) and group for time series components (gftsc) techniques
Commonly in time series modelling, identifying the four time series components which are trend, seasonal, cyclical, and irregular is conducted manually using the time series plot. However, this manual identification approach requires tacit knowledge of the expert forecaster. Thus, an automated ident...
Saved in:
| Main Author: | |
|---|---|
| Format: | Thesis |
| Language: | eng eng |
| Published: |
2022
|
| Subjects: | |
| Online Access: | https://etd.uum.edu.my/10211/1/s901903_01.pdf https://etd.uum.edu.my/10211/2/s901903_02.pdf |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Commonly in time series modelling, identifying the four time series components which are trend, seasonal, cyclical, and irregular is conducted manually using the time series plot. However, this manual identification approach requires tacit knowledge of the expert forecaster. Thus, an automated identification approach is needed to bridge the gap between expert and end user. Previously, a technique known as Break for Additive Seasonal and Trend (BFAST) was developed to automatically identify only linear trend and seasonal components, and consider the other two (i.e., cyclical and irregular) as random. Therefore, in this study, BFAST was extended to identify all four time series components using two new techniques termed Break for Time Series Components (BFTSC) and Group for Time Series Components (GFTSC). Both techniques were developed by adding cyclical and irregular components to the previous BFAST technique. The performance of BFTSC and GFTSC were validated through simulation and empirical studies. In the simulation study, monthly and yearly data were replicated 100 times based on three sample sizes (small, medium, and large), and embedding the four time series components as the simulation conditions. Percentages of identifying the correct time series components were calculated in the simulation data. Meanwhile in the empirical study, four data sets were used by comparing the manual identification approach with the BFTSC and GFTSC automatic identification. The simulation findings indicated that BFTSC and GFTSC identified correct time series components 100% when large sample size combined with linear trend and other remaining time series components. The empirical findings also supported BFTSC and GFTSC, which performed as well as a manual identification approach for only two data sets exhibiting linear trend and other components combinations. Both techniques were not performing well in other two data sets displaying curve trend. These findings indicated that BFTSC and GFTSC automatic identification techniques are suitable for data with linear trend and require future extensions for other trends. The proposed techniques help end user in reducing time to automatically identify the time series components |
|---|