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

Authors: Hao Wang, Licheng Pan, Qingsong Wen, Jialin Yu, Zhichao Chen, Chunyuan Zheng, Xiaoxi Li, Zhixuan Chu, Chao Xu, Mingming Gong, Haoxuan Li, Yuan Lu, Zhouchen Lin, Philip H. S. Torr, Yan Liu Date: 2026-03-20 Paper ID: openalex:2603.19899

Summary

This paper presents a comprehensive survey on deep time-series forecasting, framed specifically around the modeling of autocorrelation, which is a fundamental property of time-series data. The authors address two core challenges: designing architectures to capture autocorrelation in input history sequences and devising learning objectives to model autocorrelation in label sequences. The work introduces a novel taxonomy that explicitly covers both architectural design and learning objectives, unlike prior surveys that often neglect the latter. By adopting this unified, autocorrelation-centric perspective, the paper provides a deep analysis of the progression and insights driving recent advancements in the field.

Key Contributions

• Proposes a novel taxonomy for deep time-series forecasting that systematically covers both neural architectures for history autocorrelation and learning objectives for label autocorrelation.
• Provides a unified, autocorrelation-centric analysis of the progression and motivations within recent deep time-series forecasting literature, which was previously lacking in comprehensive surveys.
• Offers a holistic overview of the evolution of deep time-series forecasting by framing model design and learning strategies around the central concept of autocorrelation.

Limitations

The survey focuses primarily on recent literature and may not cover all historical or highly specialized methods outside the deep learning paradigm for time-series.

Open Questions & Future Work

• decorrelation-strategies-lightweight-models
• foundation-models-autocorrelation-tokenization
• modernizing-statistical-objectives
• advanced-generative-forecasting

Limitations

The survey focuses primarily on recent literature and may not cover all historical or highly specialized methods outside the deep learning paradigm for time-series.

Links

• Abstract
• PDF