A Survey of Deep Time Series Forecasting Backbone Architectures: Progress, Pitfalls, and a Systematic Comparison
Xiang LI , Yanping ZHENG , Zhewei WEI
Deep learning-based time series forecasting has largely converged to standardized training protocols and a narrow set of public benchmarks. While such standardization improves comparability, evaluations based on averaged metrics over fixed windows obscure variable-level differences, mask long-horizon degradation, and diverge from real-world rolling-forecasting scenarios. This study revisits these limitations by surveying seven major backbone architectures and systematically evaluating 21 representative models across diverse datasets. A fine-grained variable-level analysis shows that, in several settings, extending the input window contributes more to forecasting accuracy than architectural innovations, yet such gains diminish rapidly and may even become detrimental as the window grows excessively. Furthermore, models with strong average performance often behave inconsistently across variables and differ markedly in their ability to capture short- and long-term temporal dynamics. These findings highlight inherent constraints on predictability and call for new research directions, including domain-specific end-to-end forecasting pipelines, forecasting with exogenous drivers, and the development of large time series models.
Time series forecasting / Backbone architectures / Variable-level analysis
Higher Education Press 2026
/
| 〈 |
|
〉 |
PDF
(2561KB)
