The article describes the design of a series of competitions the aim of which is the evaluation and the further development of the state-of-the art in spatio-temporal forecasting. The means of doing so is to provide novel test data incrementally, while evaluating work of competing teams that submit algorithms in terms of performance criteria which include accuracy of predictions and time of computation. Initial results are presented hereing, whereas final results of the ongoing challenge will be presented at ICLR. 1 BACKGROUND OF THE SEE4C PROJECT We have undertaken a novel competition design to address current challenges in spatio-temporal forecasting, within the realm of an ongoing projected named SEE.4C (see-foresee) sponsored by the European commission (H2020). Spatio-temporal data (such as video) is rich and complex and it still remains an open question what type of learning algorithms (deep learning, kernel based approaches, traditional economics and engineering based approaches) provide superior performance. The challenges we address have both a long history as forecasting competitions (e.g. “M” series of challenges, the WxChallenge), in which the impetus is on repeated out-of-sample, causal regression and as ML competitions (AutoML challenge) in which the evaluation of submissions by challenge participants is not based on predictions or labels but actually on code submitted, which must run efficiently: in the case of forecasting real-time operation is necessary. The current work describes the process of evaluating algorithms for difficult learning tasks, allowing for collaborative effort and rapid progress, the goal of which is to establish a procedure and platform for evaluating algorithms that predict unseen samples adaptively, future work involves the expansion of the base of the codebase to allow for more complex state-of-the art methods applicable not only to forecasting of video