Prediction of two dimensional hand trajectories from cortical surface recordings entails finding a functional mapping from spectral modulations in multidimensional channels to instantaneous hand positions. Such studies thus far have been conducted through linear adaptive filters, even though, the functional mapping from the cortical activity to behavior might be nonlinear. Herein, we employ a nonlinear adaptive filter, kernel least mean square (KLMS), which nonlinearly map inputs to a higher dimensional feature space in which inner products can be efficiently computed. The methodology is a simple and effective nonlinear extension of the least mean square (LMS) algorithm. Preliminary results show significant improvements in mean squared error (MSE) values of reconstructed trajectories compared to linear methods (LMS) at a confidence level of 95% in the axis of highest excursion.