比神经网络好的预测算法

pluto

比神经网络好的预测算法

来自：MACD论坛(bbs.macd.cn) 作者：pluto 浏览：5424 回复：12

我用支持向量机预测股市大盘走势：
算法和神经网络比较结果：
支持向量机和神经网络都有很强的学习能力；支持向量机的泛化能力明显强于神经网络，而且很容易在拟合程度与泛化水平上进行平衡；支持向量机SMO的收敛速度要明显高于共轭梯度的神经网络，训练时间几乎提高了3个数量级；给定初始参数，支持向量机训练每次都可以得到相同的结果，这说明支持向量机得到的是全局最优解。神经网络在相同给定下，每次训练会有一定诧异结果，这也是局部极小的体现。

但我用支持向量机预测股市大盘走势，效果却没有预料的好，我想主要是输入选的不很合适，请指点。
ssbbqq@sohu.com

eagle

Support Vector Machines(SVMs) are superiror to neural networks in the sense that they have a global minimum solution once C (regulation parameter) and kernel parameters are given. So overtraining and early stopping etc is no problem for SVM. However, optimal values for these hyperparameters still rely on user's choice, although some model selection techniques can apply. You can use SVM for both regression and classification.
Stock trends, a kind of time-series data,  are qutie stochastic and instable. Noise does exist among the main trends are is not negligible. No matther how nice the tools you have, the prediction capability is quite limited and even in doubt.

pluto

May i share your opinion on ANN?
If ANN can work well on stock market forecast,why SVR or SVM not?
In fact,i've been working for a few month  and really have got some progress on this.

eagle

actually SVM can replace ANN in my opinion.
i don't know how do you use SVM in your application. what's the target you are trying to predict? the extent of advance/decline ( regression) or whether advance or decline(classification) ? what is the raw data? (index or individual shares) do you perform preprocessing on the raw data(de-noise, calculate technical indicators etc)? how do you evaluate the performance? you know, the model trained for one year may lose significant predictive power in another year? you might do some kind of these tests to have a better general idea. i understand this is quite a time-consuming task.

i2e

哪里有相信介绍？

RusselHarvey

有没有网站介绍.  听着不错.

胡逐鹿

http://www.support-vector.net/
http://www.kernel-machines.org/
http://citeseer.nj.nec.com/burges98tutorial.html
http://www.support-vector.ws/

piggreatpig

著名神经网络软件DEEPINSIGHT，能够较为精确的 给出中国股票买卖时机，有家大陆著名软件实际就是DEEPINSIGHT的变种。

野狐禅

Originally posted by pluto但我用支持向量机预测股市大盘走势，效果却没有预料的好，我想主要是输入选的不很合适，请指点。

SVM forms a quadratic problem so that it has a global optimum solution. But this is not the global optimum solution for the real problem.

For your situation, most likely you are over fitting data. You might want to reduce the vectors used in SVM until the training results similar to the results from un-seen validation data.

wym2003

Originally posted by piggreatpig at 2003-10-23 20:02
著名神经网络软件DEEPINSIGHT，能够较为精确的 给出中国股票买卖时机，有家大陆著名软件实际就是DEEPINSIGHT的变种。

谁？？？

王寰

多交流

myimage

谁说神经网络每次训练结果不同？？那是你引入了不确定性的预处理获得优化方式了了吧。支持向量机是全局最优？？

hchgsh

好！顶！