National Technical University of Athens (NTUA)
Ανάπτυξη πλαισίου αυτοματοποιημένης προέκτασης χρονοσειρών μέσω της γενίκευσης της μεθόδου πρόβλεψης θ
Abstract
dc:descriptionThe current doctoral thesis focuses on the development of an integrated forecasting methodology, able to extrapolate time series in a more efficient way. The main aim of the methodology is therefore to contribute towards automatically predicting large volume of series of various characteristics and types in a consolidated way, leading simultaneously to a robust forecasting process of high accuracy and low computational cost. In order to achieve such an objective, one must first identify and select the methods which systematically improve forecasting performance and combine them appropriately to form an integrated methodological framework. The literature indicates a variety of forecasting models and techniques for enchanting forecasting accuracy. The pros and cons of many approaches were examined, the best ones were exploited, while new ones were proposed. More specifically, the reasons which make the Theta method so effective in time series extrapolation were first detected and some expansions were suggested to improve it. This lead to an expanded version of the method which enables us to accurately forecast time series of no-linear trend patterns, as well as to combine the components of trend and level either additively or multiplicatively. The proposed model generalizes in fact the classic approach of the method by introducing multiple models of different principles. Thus, if the best model is selected per case to extrapolate a series, the average performance of the method is expected to be significantly improved. To further improve the performance of the proposed forecasting method, a new technique for smoothing time series was introduced. The technique is based on a non-parametric approach for mitigating the effect of the noise component of the data which enables us to extract the real signal of the series and extrapolate it. This is of vital importance since, in contrast to typical approaches reported in the literature, the technique exploits a nonlinear algorithm to shrink local variances and curvatures leading to a far more smoothed series. The smoothed series can then be easily extrapolated by any forecasting model, leading to better forecasts. The generalized form of the Theta method is combined with the smoothing technique introduced and many other algorithms found in the literature to develop a complete forecasting methodology. The effectiveness of the methodology is evaluated using both real and generated data. The real data refer to the time series of the M competitions, while the generated data are simulated using a process which was developed for this purpose. That way the conclusions extracted are representative and based on multiple series of various characteristics. The results of the thesis are more than encouraging, reporting the superiority of the proposed approach.
Degree
thesis:*- Grantor dc:publisher
- National Technical University of Athens (NTUA)
- Year dc:date
- 2017
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Spiliotis, Evangelos
- Σπηλιώτης, Ευάγγελος
Subjects
dc:subject × 14- Ακρίβεια πρόβλεψης
- Εξομάλυνση χρονοσειρών
- Μέθοδος Θ
- Πρόβλεψη χρονοσειρών
- Τεχνικές προβλέψεων
- Forecasting accuracy
- Smoothing
- Theta method
- Time series forecasting
- Forecasting techniques
- Επιστήμες Μηχανικού και Τεχνολογία
- Επιστήμη Ηλεκτρολόγου Μηχανικού, Ηλεκτρονικού Μηχανικού, Μηχανικού Η/Υ
- Engineering and Technology
- Electrical Engineering, Electronic Engineering, Information Engineering
Rights
- Language dc:language
- gre
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/41725