In this paper we report an experimental evaluation of three popular methods for online system identification of unmanned surface vehicles (USVs) which were implemented as an ensemble: certifiably stable shallow recurrent neural network (RNN), adaptive identification (AID), and recursive least squares (RLS). The algorithms were deployed on eight USVs for a total of 30 hours of online estimation. During online training the loss function for the RNN was augmented to include a cost for violating a sufficient condition for the RNN to be stable in the sense of contraction stability. Additionally we described an efficient method to calculate the equilibrium points of the RNN and classify the associated stability properties about these points. We found the AID method had lowest mean absolute error in the online prediction setting, but a weighted ensemble had lower error in offline processing.