Kolmogorov-Arnold Networks (KAN) are a novel neural network architecture inspired by the Kolmogorov-Arnold representation theorem, offering a promising alternative to traditional models like MLPs. This paper explores the application of KAN to time series forecasting, introducing two variants: T-KAN and MT-KAN. T-KAN is designed to detect and track concept drift in univariate time series, using symbolic regression to enhance interpretability. MT-KAN extends this to multivariate time series, effectively capturing complex variable interactions. Both models outperform traditional methods in forecasting tasks, achieving high accuracy and interpretability. KAN's use of spline-parametrized univariate functions allows for dynamic learning and adaptability, improving both performance and transparency. The paper also discusses the limitations of KAN, including slower training times compared to MLPs, and suggests potential future directions for integrating KAN with other architectures to enhance flexibility and efficiency. The results show that KAN-based models can achieve high accuracy with fewer parameters, making them suitable for real-time applications. The study highlights the potential of KAN as a powerful and interpretable tool in predictive analytics.Kolmogorov-Arnold Networks (KAN) are a novel neural network architecture inspired by the Kolmogorov-Arnold representation theorem, offering a promising alternative to traditional models like MLPs. This paper explores the application of KAN to time series forecasting, introducing two variants: T-KAN and MT-KAN. T-KAN is designed to detect and track concept drift in univariate time series, using symbolic regression to enhance interpretability. MT-KAN extends this to multivariate time series, effectively capturing complex variable interactions. Both models outperform traditional methods in forecasting tasks, achieving high accuracy and interpretability. KAN's use of spline-parametrized univariate functions allows for dynamic learning and adaptability, improving both performance and transparency. The paper also discusses the limitations of KAN, including slower training times compared to MLPs, and suggests potential future directions for integrating KAN with other architectures to enhance flexibility and efficiency. The results show that KAN-based models can achieve high accuracy with fewer parameters, making them suitable for real-time applications. The study highlights the potential of KAN as a powerful and interpretable tool in predictive analytics.