DeepEDM: A New Framework for Nonlinear Time Series Forecasting
Researchers have introduced DeepEDM, a framework that combines nonlinear dynamical systems modeling with deep neural networks for time series forecasting. Inspired by empirical dynamic modeling (EDM) and Takens' theorem, DeepEDM learns a latent space from time-delayed embeddings and uses kernel regression to approximate underlying dynamics. It leverages efficient softmax attention for accurate multi-step predictions. The method was evaluated on synthetic nonlinear dynamical systems and real-world time series across various domains.
Key facts
- DeepEDM integrates nonlinear dynamical systems modeling with deep neural networks.
- It is inspired by empirical dynamic modeling (EDM) and Takens' theorem.
- The framework learns a latent space from time-delayed embeddings.
- It uses kernel regression to approximate underlying dynamics.
- Softmax attention is employed for efficient implementation.
- Experiments were conducted on synthetic and real-world time series.
- The paper is available on arXiv with ID 2506.06454.
- The approach aims to improve prediction by explicitly modeling dynamics.
Entities
Institutions
- arXiv