Abstract: Although the amount of massive time series data or data flow has continued to increase, the classification training speed of the twin parametric-margin support vector machine (TPMSVM) is still very slow. This study provides proof of the sample satisfying the KKT condition of the TPMSVM corresponding to the numerical condition. According to the conclusion, an incremental learning algorithm enabling the TPMSVM to deal with time series data is proposed. This algorithm selects the new data that violate the generalized KKT condition and the partial original data that satisfies the condition to participate in the classifier training. The experimental results show that the proposed algorithm not only maintains a certain classification accuracy but can also improve the accuracy of the TPMSVM training.