Time series prediction based on sparse SVR

Sparse support vector regression (SpSVR) method is proposed to improve the leaning speed without decreasing generalization performance. Firstly, the primal problem of support vector regression is directly optimized through Newton optimization method. Then, in order to realize the sparseness of SVR, Cholesky decomposition is used to update the Hessian matrix in SVR primal problem. Finally, such proposed algorithm is applied into Mackey-Glass, Lorenz and Logistic chaotic time series prediction. The simulation results indicate that the SpSVR is able to effectively reduce the number of support vectors with guaranteed prediction precision.