In this paper we study the downlink of a TDD (Time Division Duplex) single cell system where the Base Station (BS) employs multiple antennas to serve the users taking into account the traffic patterns. The BS chooses each slot the users to be active, and serves them using Zero Forcing (ZF) precoding. This requires the knowledge of the users’ channels which is assumed to be performed e.g. via uplink training. Due to the channel acquisition overhead, only a subset of users must be active at each timeslot (depending on traffic patterns and channel states). In this paper, we develop an active user selection strategy where the base station sets a given threshold for the channel gain of the users. Then, only the users that have their channel gain higher than the threshold send their training sequence and are then considered to be served. The base station estimates the channel states of these users and, due to channel reciprocity, uses these channel states to transmit data via ZF precoding. With appropriate signaling and threshold selection, which adapt to the queuing behavior of the users, we prove that our proposed method achieves a larger stability region than the baseline centralized policy where the BS selects the users based on channel statistics and queue lengths. The performance of the threshold-based method is illustrated via simulations, where we can observe a tradeoff between the expansion of the stability region and delay performance.