Abstract
This paper introduces a novel application of Kolmogorov-Arnold Networks(KANs) to time series forecasting, leveraging their adaptive activationfunctions for enhanced predictive modeling. Inspired by the Kolmogorov-Arnoldrepresentation theorem, KANs replace traditional linear weights withspline-parametrized univariate functions, allowing them to learn activationpatterns dynamically. We demonstrate that KANs outperforms conventionalMulti-Layer Perceptrons (MLPs) in a real-world satellite traffic forecastingtask, providing more accurate results with considerably fewer number oflearnable parameters. We also provide an ablation study of KAN-specificparameters impact on performance. The proposed approach opens new avenues foradaptive forecasting models, emphasizing the potential of KANs as a powerfultool in predictive analytics.