Forecasting keyword/topic activities in social networking sites has been the subject of many recent studies, as such activities represent, in many cases, a direct estimate of the spreadth of real-world phenomena, ıt eg. box-office revenues or flu epidemies. Most of these studies rely on pointwise, regression-like prediction algorithms and focus on few, usually unambiguous, keywords/topics. We study different strategies to rank keyword activities through a comparison of pointwise and pairwise learning to rank approaches, as well as the impact of keyword ambiguity, keyword activity and keyword set size on the prediction results. It is the first time, to our knowledge, that such dimensions are evaluated in this framework. Our experiments are conducted on a large dataset built by monitoring Twitter over a year and including 1497 keywords.