This paper addresses the challenges of irregular sampling intervals and missing values in real-world time series data, which conventional methods struggle with. Neural Ordinary Differential Equations (Neural ODEs) and Neural Stochastic Differential Equations (Neural SDEs) are proposed as alternative approaches to learn continuous latent representations. Neural SDEs extend Neural ODEs by incorporating a diffusion term, but this addition is not straightforward, especially when dealing with irregular intervals and missing values. The authors propose three stable classes of Neural SDEs: Langevin-type SDE, Linear Noise SDE, and Geometric SDE. These models are designed to maintain stability and enhance performance while preventing overfitting under distribution shift. Extensive experiments on benchmark datasets for interpolation, forecasting, and classification tasks demonstrate the effectiveness of the proposed methods, showing superior performance and robustness to missing data. The theoretical analysis provides insights into the existence and uniqueness of solutions, as well as the robustness of the models under distribution shift.This paper addresses the challenges of irregular sampling intervals and missing values in real-world time series data, which conventional methods struggle with. Neural Ordinary Differential Equations (Neural ODEs) and Neural Stochastic Differential Equations (Neural SDEs) are proposed as alternative approaches to learn continuous latent representations. Neural SDEs extend Neural ODEs by incorporating a diffusion term, but this addition is not straightforward, especially when dealing with irregular intervals and missing values. The authors propose three stable classes of Neural SDEs: Langevin-type SDE, Linear Noise SDE, and Geometric SDE. These models are designed to maintain stability and enhance performance while preventing overfitting under distribution shift. Extensive experiments on benchmark datasets for interpolation, forecasting, and classification tasks demonstrate the effectiveness of the proposed methods, showing superior performance and robustness to missing data. The theoretical analysis provides insights into the existence and uniqueness of solutions, as well as the robustness of the models under distribution shift.