Deep Autocorrelation Modeling for Time-Series Forecasting: Progress and Prospects

A research consortium of 15 authors from institutions including Tsinghua University, University of Oxford, and Monash University has published a landmark survey titled 'Deep Autocorrelation Modeling for Time-Series Forecasting: Progress and Prospects' on arXiv. The paper addresses a critical gap in AI literature by providing the first comprehensive review that systematically examines both neural architectures and learning objectives through the lens of autocorrelation—the statistical dependency between sequential observations that defines time-series data. Unlike previous surveys that focused primarily on model structures, this work offers a unified framework for understanding 15 years of progress in deep forecasting.

The survey makes two distinctive contributions that will shape future research. First, it introduces a novel taxonomy that categorizes existing literature across both architectural innovations (how models capture historical patterns) and objective functions (how they learn from future sequences). Second, it provides thorough analysis of the motivations, insights, and progression of surveyed methods from a consistent autocorrelation perspective. The authors have compiled extensive resources including a full list of papers and tools, creating what will likely become the definitive reference for researchers working on forecasting applications from finance to climate science to industrial IoT.