改进SVM及其在时间序列数据预测中的应用

Modified SVM and Its Application to Time Series Forecasting

运用标准支持向量机预测海量金融时间序列数据会出现训练速度慢、内存开销大的问题,文中提出一种分解合作加权的回归支持向量机,将大样本集分解成若干工作子集,分段提炼出支持向量机,同时根据支持向量的重要性给出不同的错误惩罚度,并将其应用于证券指数预测.与标准算法相比较,文中方法在保证泛化精度一致的前提下,极大地加快了训练速度.

As the forecasting of a huge-size financial time series by training a standard SVM (Support Vector Machine) will result in slow training speed and large memory spending, a decomposition-cooperation-weighted SVM regression is put forward and used to predict the stock index. In the proposed method, a large specimen set is decomposed into several subsets, and the SVMs in different subsets are independently extracted. According to the importance of the obtained SVMs, different error punishment degrees are obtained. Compared with the traditional SVM, the proposed method greatly speeds up the training process with almost the same precision.