Experimental applications of nonlinear system identification algorithms to a reaction wheel pendulum are presented in this paper. In particular, the subspace instrumental variable identification algorithm proposed in [D. Luo and A. Leonessa, 2002] is applied to a nonlinear prediction model represented by a Hammerstein system with nonlinear feedbacks. The prediction performances are compared with the experimental results by performing an open loop subspace identification on a linear state-space prediction model. Some practical aspects on the nonlinear identification are discussed. Experimental results show the effectiveness of the SIVID algorithm on system model recovery.