Bike-sharing systems are widely deployed in many major cities, providing a convenient transportation mode for citizens’ commutes. As the rents/returns of bikes at different stations during different periods are unbalanced, the bikes in a system need to be rebalanced all the time. Real-time monitoring cannot tackle this problem well as it is too late to reallocate bikes after an unbalance has occurred. In this paper, we propose a hierarchical prediction model to predict the check-out/in of each station cluster in a future period so that reallocation can be executed in advance. We propose a bipartite clustering algorithm to cluster stations into groups, based on which the hierarchical prediction of check-out/in can be done. After the entire traffic of the whole city is obtained by a Gradient Boosting Regression Tree (GBRT), a multi-similarity-based inference model is proposed to predict the check-out proportion across clusters and the inter-cluster transition matrix. Thus, the check-out across clusters can be calculated easily. Then, each cluster’s check-in is inferred from the check-out, inter-cluster transition and trip duration. We evaluate our models on two bike-sharing systems in New York (NY) and Washington (WA), respectively, to confirm our models’ advantages beyond baseline approaches.