This work describes the application of nonlinear kernel based regression method for prediction of 3 degrees of freedom (DoF) wrist movement from surface electromyogram (sEMG) signals. Here, two semicircular motion trajectories have been used, where multiple DoFs are engaged in a free wrist movement and near the both extremities of the motion profile all the three DoFs of wrist are involved. Compared to earlier studies this is novel in the fact that the motion profiles are with free and unconstrained wrist movement and are closely related to the real life complex movement situations. The proposed method involves learning the wrist motion profile of a subject from a known data set comprising of motion sensor, sEMG and then tries to predict the motion profile from a new and unknown dataset of sEMG only from the same subject. Prediction accuracies for data in cross validation stage and for new data are presented. Though the accuracy of prediction is poorer in the latter case since we are considering conditions of free hand movements as in practical situations the result shown is quite significant considering the focus of developing independent, proportional and simultaneous control of multiple DoFs of prosthesis.