Bagging ensemble-based novel data generation method for univariate time series forecasting. (1st October 2022)

• Record Type: Journal Article
• Authors: Kim, Donghwan; Baek, Jun-Geol
• Is Part Of: Expert systems with applications. Volume 203(2022)
• Publication Date: 2022-10-01
• Subjects: Time series forecasting -- Ensemble method -- Bagging -- Neural network -- Maximum overlap discrete wavelet transform -- Data augmentation; Expert systems (Computer science) -- Periodicals; Systèmes experts (Informatique) -- Périodiques; Electronic journals; 006.33
• DOI: 10.1016/j.eswa.2022.117366
• Languages: English
• ISSNs: 0957-4174

Abstract

Highlights: We propose a bagging based data generation method for univariate time series forecasting. We combine a wavelet transform with bootstrap for generating an ensemble model. The proposed method effectively enables univariate time-series data to the ensemble. The proposed method overcomes comparative prediction methods. The utility of the proposed method was proved through additional experiments. Abstract: The most critical issue in time series data is predicting future data values. Recently, an ensemble model combining multiple models with superior predictive performance has emerged. However, in the case of univariate time series data, an accurate prediction remains difficult because of the unique characteristic of the data: there is only one variable to analyze. In this paper, we propose a method to improve the performance of predictive models with a simple structure and apply it to time series data. This study proposes a time series forecasting method based on a bagging ensemble that uses the maximum overlap discrete wavelet transform (MODWT) and bootstrap. The proposed method decomposes the scale and detail of the time series data using the MODWT. The bootstrap is applied to univariate time series to generate bootstrapped data that slightly differ from the characteristics of the original data. Through experiments, we examined the results and validated the details of the proposed method depending on whether the proposed method was applied. In most cases, we confirmed that our proposed method improves the performance of the existing algorithms by employing a nonparametric test. The results show that the performance improved more when the algorithm is simple.