摘要
为了解决神经网络算法预测海量金融时间序列数据会出现训练速度慢,内存开销大等问题,提出一种基于最小二乘支持向量机的复杂金融数据时间序列预测方法。该方法将传统的支持向量机中的不等式约束改为等式约束,且将误差平方和的损失函数作为训练集的经验函数,这样把二次规划问题转化为求解线性方程组问题,提高求解问题的速度和收敛精度。实验中以证券指数为实验数据,对大批量金融数据进行了时间序列预测,相比于神经网络预测方法,该方法在大批量金融数据时间序列预测的训练时间、训练次数和预测误差上都有了明显提高,对复杂金融时间序列具有较好的预测效果。
Predictions of massive financial time series data using neural networks calculation bring problems such as slow training speeds and large memory cost. This paper describes a complex financial data time series prediction method based on the least square support vector machine （LS-SVM）, which changes the inequality restriction in the traditional SVM into an equality restriction and uses the loss function of the quadratic sum of the errors as the empirical function for the training set. In this way, the quadratic programming problem is converted into solving linear equations, which significantly improves the solving speed and the convergence accuracy. As an example, financial time series data for the stock index was well predicted by the method with the faster training time and better accuracy than neural network forecasts.
出处
《清华大学学报（自然科学版）》
EI
CAS
CSCD
北大核心
2008年第7期1147-1149,共3页
Journal of Tsinghua University(Science and Technology)
基金
国家"九七三"重点基础研究项目(2004CB719406)
国家"八六三"高技术项目(2004AA413010)
关键词
金融数据
时间序列预测
支持向量机
financial data
time series prediction
support vector machine
