Methods and apparatuses to model the relation between the delay samples and congestion losses and to protect media flows against congestion losses are disclosed. In one embodiment, the method comprises measuring congestion by delay samples, and performing a dynamic FEC algorithm that uses convex hull clustering for loss-event classification, including determining an FEC rate according to the loss-event classification. In another embodiment, the method comprises measuring congestion by delay samples, modeling loss events associated with the delay samples by grouping loss events as unions of convex hulls to identify a period of potentially increased congestion, dynamically changing the FEC rate based on the modeling, and applying FEC protection to the media flow during the period based on the FEC rate. In yet another embodiment, the method comprises forming clusters using past loss events and their {FTT, F{dot over (T)}T} or {RTT, R{dot over (T)}T} estimates, and estimating future loss events using most recent {FTT, F{dot over (T)}T} or {RTT, R{dot over (T)}T} estimates and their membership in convex hull clusters.