A dilemma in cloud-radio access networks (C-RANs) is to balance the cluster scale and the cooperative gains. In this paper, cluster formation for downlink transmissions in C-RANs is studied. In particular, with a fixed intro-cluster cooperation strategy, an explicit expression for the successful access probability is derived by applying stochastic geometry. By using the obtained theoretical result as a utility function, the problem of grouping remote radio heads is formulated as a coalitional formation game, and then two distributed algorithms based on the merge and split approach are obtained as efficient solutions for the cases with and without cluster size constraints, respectively. Compared with grand cluster formation, which is a centralized method, simulation results show that the proposed approaches can achieve better performance with smaller cluster settings.