Revisiting, Rethinking, and Benchmarking Deep Time Series Forecasting

The Hong Kong University of Science and Technology (Guangzhou)

Data Science and Analytics Thrust

PhD Qualifying Examination

By Ms. Jiawen ZHANG

Abstract

The ubiquitous proliferation of sensor technologies and digitization processes has rendered time series data an indispensable asset across numerous domains. This data, methodically collected and indexed over time, plays a pivotal role in diverse analytical strategies employed in fields ranging from economic forecasting to scientific research and engineering. This survey provides a comprehensive examination of the evolution and current trends in deep time series forecasting, contrasting traditional methodologies with modern deep learning approaches. We deconstruct current deep learning methods into three key dimensions: neural network architecture, probabilistic estimation techniques, and training strategies. Significant gaps in existing research are identified, and in response, we introduce ProbTS—a robust research toolkit crafted to benchmark, analyze, and evaluate various forecasting models. Through extensive experimentation, this tool reveals intriguing insights and highlights promising research directions. By offering a thorough literature review and empirical analysis, this survey aims to inspire and propel future research in time series forecasting.

PQE Committee

Chairperson: Prof. Fugee TSUNG

Prime Supervisor: Prof Jia LI

Co-Supervisor: Prof Xiaofang ZHOU

Examiner: Prof Yuyu LUO