Automatic selection of the best neural architecture for time series forecasting
Automatic selection of the best neural architecture for time series forecasting
Authors: Qianying Cao, Shanqing Liu, Alan John Varghese, Jérôme Darbon, Michael S. Triantafyllou, George Em Karniadakis Date: 2026-06-02 Paper ID: openalex:2501.12215
Summary
This paper addresses the challenge of selecting an optimal neural architecture for time series forecasting by proposing an automated framework that constructs and evaluates hybrid models. By employing multi-objective optimization to explore combinations of LSTM, GRU, attention, and state-space model blocks, the framework identifies Pareto-optimal designs tailored to specific application objectives. Additionally, the authors introduce two sampling-based iterative techniques that significantly accelerate the search process, reducing total training costs by nearly eightfold. The results indicate that no single architecture is universally superior, advocating for data-driven, objective-specific architecture search in forecasting tasks.
Key Contributions
- Proposes an automated framework for designing hybrid time series forecasting models by combining LSTM, GRU, attention, and SSM modules.
- Introduces two sampling-based iterative procedures for Pareto-front exploration that reduce total training cost by nearly eightfold.
- Demonstrates that hybrid architectures outperform monolithic designs when balancing accuracy and complexity requirements across diverse benchmarks.
Links
Metadata & Links
- url
- https://arxiv.org/abs/2501.12215
- paper_id
- 2501.12215
- paper_source
- openalex
- domain
- time-series
- tags
-
- architectures
-
- datasets
-
- concept_slugs
-
- dataset_slugs
-
- skill
- TimeSeriesSkill
- created_at
- 2026-06-05T06:31:52Z