Mass events represent one of the most challenging scenarios for mobile networks because, although their date and time are usually known in advance, the actual demand for resources is difficult to predict due to its dependency on many different factors. Based on data provided by a major European carrier during mass events in a football stadium comprising up to 30.000 people, 16 base station sectors and $1$Km$^2$ area, we performed a data-driven analysis of the radio access network infrastructure dynamics during such events. Given the insights obtained from the analysis, we developed ARENA, a model-free deep learning Radio Access Network (RAN) capacity forecasting solution that, taking as input past network monitoring data and events context information, provides guidance to mobile operators on the expected RAN capacity needed during a future event. Our results, validated against real events contained in the dataset, illustrate the effectiveness of our proposed solution.