This paper presents a novel method for enhancing the expressive power of Neural Coefficients Derived (CDEs), a natural way to process irregular time series. Existing Neural CDEs adjust the depth (the number of function evaluations, or NFE) of the model by lowering the solver error tolerance to increase numerical accuracy. However, this approach is insufficient for enhancing expressive power. In this paper, we present a simple yet effective alternative: expanding the integration time horizon to increase NFE and deepen the model. To address the uncontrolled growth of the original vector field, we propose the Negative Feedback (NF) technique, which guarantees stability without restricting flexibility. We provide theoretical bounds on the risk of Neural ODEs using Gaussian process theory, and experiments on four public datasets demonstrate that the proposed method outperforms existing methods, including Neural RDEs and state-space models, by up to 20%. The proposed method, called DeNOTS, combines expressive power, stability, and robustness to enable reliable modeling in the continuous-time domain.
