Wavelets meet transformers: an experimental approach to time series forecasting

• Main Author: Ng, Andrew Yong Kuan
• Other Authors: Yeo Chai Kiat
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2024
• Subjects: Computer and Information Science
• Online Access: https://hdl.handle.net/10356/175093
• Institution: Nanyang Technological University

Forecasting time series data accurately is paramount across a spectrum of disciplines ranging from finance to environmental science. In recent years, the application of advanced machine learning techniques, particularly deep learning models, has shown promising results in this field. This report delves into the innovative integration of stationary wavelet transformation (SWT) and transformer-based models to enhance time series data prediction accuracy. The study systematically evaluates the performance of a hybrid model that combines SWT and transformers. Firstly, SWT is employed as a preprocessing step to decompose the time series data into different frequency components. Subsequently, these components serve as input to a transformer-based model designed to capture complex temporal dependencies. Through empirical analysis of several benchmark time series datasets, this report aims to demonstrate that the hybrid approach outperforms traditional methods and standalone transformer models in terms of prediction accuracy. Furthermore, the adaptability of the hybrid model opens avenues for the exploration of other wavelet transformations. This study proposes the Maximal Overlap Discrete Wavelet Transform (MODWT) as a viable alternative to address the limitations encountered with the Stationary Wavelet Transformation (SWT).