The stock market reflects the development of the national economy. More and more people and institutions put their money into the stock market. At present, the stock market urgently needs an efficient investment theory. An improved SVR algorithm is proposed for the case that the prediction error of SVR to individual stocks is very large in this paper. The improved algorithm is simulated by using the most recent data as the training data instead of all the historical data. The improved algorithm was compared with SVR and decision tree algorithms using decision coefficients and root mean square as evaluation indicators. The experimental results show that the improved SVR Algorithm is obviously superior to SVR.