This paper presents a model-free predictor applied to forecast values of signal-to-noise ratio (SNR) relation in wireless communication networks. The main motivation behind this study is to improve the spectrum allocation in 5 G networks by proactively selecting a proper modulation and coding scheme using predicted (SNR) values rather than relying on real-time measurements. To achieve this, the proposed method combines techniques from chaos theory, specifically Takens reconstruction, with the unscented Kalman filter to generate a local model that is used to estimate the evolution of the system. This model-free predictor is called the Kalman-Takens filter (KTF). When applied to data collected from a live 5 G the error between predicted and observed SNRs were mostly below 10% after the filter had enough time to converge. The optimal prediction window is around 100 time steps and no significant deviations were oberved for forecasts up to 500 steps making the KTF a useful tool for optimizing resource allocation in mobile networks.