This paper introduces the KARMA model, which integrates Kolmogorov-Arnold Networks (KANs) with Recurrent Neural Networks (RNNs) and the ARIMA framework to enhance time series forecasting (TSF) capabilities. Establishing a KAN baseline is crucial as it provides a structured approach to function approximation, allowing for improved modeling of non-linearities often present in real-world data. KANs, founded on the Kolmogorov-Arnold Theorem, have demonstrated resilience against scarce data, making them particularly effective in scenarios where traditional models struggle. By combining KANs with RNNs, the KARMA model not only leverages the strengths of both architectures but also addresses the limitations of existing methods, thereby setting a benchmark for future research in TSF. The results indicate that KAN-based models consistently outperform RNNs in univariate forecasting, particularly in datasets characterized by high autocorrelation.