Streaming data analysis presents numerous challenges: the unbounded and sequential nature of the stream requires evolving models in order to capture its intrinsic non-stationarity and related concept drift. Relatively few methods for the detection of the drift are unsupervised and require no labels during training. In this paper, a soft and dynamic metaclustering approach is used for concept drift detection on streaming data. The pro posed method exploits possibilistic memberships for a natural quantifica tion of the drift and is able to effectively recognize both abrupt and recur ring drift. The effectiveness of the proposed method has been assessed on synthetic data from a public benchmark, yielding notable results.