This paper introduces Kolmogorov-Arnold Networks (KANs) for time series forecasting, demonstrating their superior performance compared to traditional Multi-Layer Perceptrons (MLPs). KANs leverage the Kolmogorov-Arnold representation theorem, replacing linear weights with spline-parametrized univariate functions, enabling dynamic learning of activation patterns. The study shows that KANs achieve more accurate results with fewer parameters in a real-world satellite traffic forecasting task. An ablation study highlights the impact of KAN-specific parameters on performance. The paper presents a novel application of KANs to time series forecasting, analyzing their efficiency in terms of trainable parameters and forecasting performance. Real-world satellite traffic data is used to evaluate KANs, showing their effectiveness in capturing complex temporal patterns. KANs outperform MLPs in terms of error metrics, with lower Mean Squared Error (MSE), Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), and Mean Absolute Percentage Error (MAPE). Additionally, KANs are more parameter-efficient, using significantly fewer parameters than MLPs. The study also explores the impact of node counts and grid sizes on KAN performance. Increasing nodes and grid sizes generally improves performance, but this comes at the cost of increased computational demands. The best performance is achieved with a high node count and large grid size, offering the highest flexibility and learning capacity. However, this may not be suitable for real-time applications due to longer training times. KANs are shown to be a promising alternative to traditional MLPs in traffic management, offering superior forecasting performance and parameter efficiency. The study emphasizes the importance of optimizing node counts and grid sizes to enhance model performance. While KANs show promise, further research is needed to optimize their use across broader applications and to develop more complex solutions that can compete with advanced architectures like LSTMs, GRUs, and CNNs.This paper introduces Kolmogorov-Arnold Networks (KANs) for time series forecasting, demonstrating their superior performance compared to traditional Multi-Layer Perceptrons (MLPs). KANs leverage the Kolmogorov-Arnold representation theorem, replacing linear weights with spline-parametrized univariate functions, enabling dynamic learning of activation patterns. The study shows that KANs achieve more accurate results with fewer parameters in a real-world satellite traffic forecasting task. An ablation study highlights the impact of KAN-specific parameters on performance. The paper presents a novel application of KANs to time series forecasting, analyzing their efficiency in terms of trainable parameters and forecasting performance. Real-world satellite traffic data is used to evaluate KANs, showing their effectiveness in capturing complex temporal patterns. KANs outperform MLPs in terms of error metrics, with lower Mean Squared Error (MSE), Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), and Mean Absolute Percentage Error (MAPE). Additionally, KANs are more parameter-efficient, using significantly fewer parameters than MLPs. The study also explores the impact of node counts and grid sizes on KAN performance. Increasing nodes and grid sizes generally improves performance, but this comes at the cost of increased computational demands. The best performance is achieved with a high node count and large grid size, offering the highest flexibility and learning capacity. However, this may not be suitable for real-time applications due to longer training times. KANs are shown to be a promising alternative to traditional MLPs in traffic management, offering superior forecasting performance and parameter efficiency. The study emphasizes the importance of optimizing node counts and grid sizes to enhance model performance. While KANs show promise, further research is needed to optimize their use across broader applications and to develop more complex solutions that can compete with advanced architectures like LSTMs, GRUs, and CNNs.