Abstract��Most chaotic time series prediction algorithms based on linear superposition of basis functions are static and lack of corresponding interpretations between the basis functions and the underlying dynamics. Improvement was made on that these functions that can approximate non-Gaussian were used as basis set, as building a relationship between the basis functions and higher-order statistics of chaotic time series. Furthermore, a nonlinear feedback was applied in the algorithm that can introduce dynamic quality. Simulations show an enhanced performance on both one step prediction error and maximum attempted time, which outperforms the linear prediction and some existing adaptive ones.