Accurate stock price predicting is a key problem to the financial field. Comparing with the traditional stock price predicting models such as GARCH models and neural networks, the theoretical advantage of applying support vector machine (SVM) to stock price predicting highly depends on solving the problem of kernel function construction and parameter optimization. For the effect of the kernel function in the SVM classification model, a hybrid kernel function is presented. In order to optimize and adjust the important parameters during the process of building the hybrid kernel function, an improved particle swarm optimization which has better global search ability is used. Experimental results about stock price index predicting show that this method has higher prediction accuracy compared with the traditional kernel functions.