Techniques for dynamic RF site configuration are provided. Historical association data is collected from a plurality of access points in a physical environment, and a machine learning model is trained to predict future association events, based on the historical association data. Current association data is then collected from the plurality of access points, and at least one predicted association event is generated by processing the current association data using the trained machine learning model. The plurality of access points is allocated to a plurality of radio frequency (RF) sites based on the at least one predicted association event. Finally, at least one of the plurality of RF sites is configured based on the predicted association event.