A New Paradigm for Time Series Modeling: A Dynamical Systems Perspective

Targeting ICML 2026, this paper presents the Dynamical Systems (DS) perspective as the direction time series modeling should move toward. Time series data in nature and engineering mostly follow complex dynamical rules, and understanding these is key to improving prediction accuracy.

The researchers propose the following four core strategies to overcome the limitations of current time series models.

• Adopting DSR-specific training techniques: Beyond simple forecasting, training objectives centered on Dynamical Systems Reconstruction (DSR), which understands the dynamical rules of a system, should be established. This allows models to capture long-term statistical properties while reducing parameter complexity.
• Pretraining on simulation-based data: Models should be pretrained on simulation data from actual dynamical systems, rather than artificial functions, in order to secure natural priors.
• A return from Transformers to modern RNNs: Instead of Transformers, which ignore temporal recursion—a core aspect of time series—the use of modern RNN architectures, which better preserve temporal structure, is recommended.
• Solving persistent challenges: An approach is needed to address the chronic problems of time series modeling, namely out-of-domain generalization and long-term behavior prediction.