The constant increase in wireless handheld devices and the prospect of billions of connected machines has compelled the research community to investigate different technologies which are able to deliver high data rates, lower latency and better reliability and quality of experience to mobile users. One of the problems, usually overlooked by the research community, is that more connected devices require proportionally more signalling overhead. Particularly, acquiring users' channel state information is necessary in order for the base station to assign frequency resources. Estimating this channel information with full resolution in frequency and in time is generally impossible, and thus, methods have to be implemented in order to reduce the overhead. In this paper, we propose a channel quality estimation method based on the concept of Gaussian process regression to predict users' channel states for varying user mobility profiles. Furthermore, we present a dual-control technique to determine which is the most appropriate prediction time for each user in order to keep the packet loss rate below a pre-defined threshold. The proposed method makes use of active learning and the exploration-exploitation paradigm, which allow the controller to choose autonomously the next sampling point in time so that the exploration of the control space is limited while still reaching an optimal performance. Extensive simulation results, carried out in an LTE-A simulator, show that the proposed channel prediction method is able to provide consistent gain, in terms of packet loss rate, for users with low and average mobility, while its efficacy is reduced for high-velocity users. The proposed dual-control technique is then applied, and its impact on the users' packet loss is analysed in a multicell network with proportional fair and maximum throughput scheduling mechanisms. Remarkably, it is shown that the presented approach allows for a reduction of the overall channel quality signalling by over 90 % while keeping the packet loss below 5 % with maximum throughput schedulers, as well as signalling reduction of 60 % with proportional fair scheduling.