In one embodiment, a device in a network obtains tunnel flappiness metrics associated with a particular tunnel in the network exhibiting flapping. The device makes, based on the tunnel flappiness metrics, a prediction that the particular tunnel is going to flap. The prediction is made using a machine learning model. The device proactively reroutes, based on the prediction, traffic from the particular tunnel onto an alternate tunnel, prior to the particular tunnel flapping. The device evaluates performance of the alternate tunnel, after proactively rerouting the traffic from the particular tunnel onto the alternate tunnel.