The tracking accuracy of the traditional online least squares support vector regression in solving regression problem of the time-varying objects is not high enough and support vectors are not sparse. To deal with this problem, an online adaptive recursive reduced least squares support vector regression is proposed by combining with the iterative strategy and reduced technique. The method considers the constrainable impact on the existing model, which is caused by the joint action of new samples and historical data. Meantime, the training sample leading to the largest reduction in the target function is chosen as the best new support vectors. Then the regression model is simplified, and the prediction time is shortened. Finally, simulation analysis illustrates the effectiveness and feasibility of the presented method. Compared with the traditional algorithms, the method is more accurate and sparse.