c*****e 发帖数: 3226 | 1 【 以下文字转载自 Stock 讨论区 】
发信人: CRH1235 (江左没狼), 信区: Stock
标 题: Re: [BSSD] GTX1080是业余GPU
发信站: BBS 未名空间站 (Thu Sep 1 16:40:33 2016, 美东)
image处理主要以convolutional network为基础,speech和股价都可以看作time
series数据,所以recurrent network比较流行。
Deep learning在金融界的应用我了解不多,不过很多用recurrent network,尤其是
LSTM做股票预测的,一搜有好多论文呢,github有很多source code可以参考。
高频的散户没搞头,要拼硬件,中长期的可以搞搞作参考,估计效果比看KDJ/MACD也好
不了多少。
Deep Learning和传统machine learning比的优势,我个人认为,是比较不要求domain
knowledge。传统machine learning都要hand crafted features,要数学功底好才行,
deep learning 只要有大量数据,建好model扔给机器算就好了(当然构建和调试model
也是很需要功力的)。我懂得也有限,暂时没想过用这个赚钱。有空了可以写个model
玩一玩。 | c*****e 发帖数: 3226 | 2 这个讲的很好:
A convolutional network is basically a standard neural network that's been
extended across space using shared weights. A recurrent neural network is
basically a standard neural network that's been extended across time by
having edges which feed into the next time step instead of into the next
layer in the same time step. There's a kind of similarity between the two
but it's pretty abstract (easier to see if you unroll the recurrent neural
network)
Intuitively, we can see that CNNs are ideal for images and videos while RNNs
are ideal for text and speech.
Although, we do find CNNs being quite useful even in text classification.
domain
【在 c*****e 的大作中提到】 : 【 以下文字转载自 Stock 讨论区 】 : 发信人: CRH1235 (江左没狼), 信区: Stock : 标 题: Re: [BSSD] GTX1080是业余GPU : 发信站: BBS 未名空间站 (Thu Sep 1 16:40:33 2016, 美东) : image处理主要以convolutional network为基础,speech和股价都可以看作time : series数据,所以recurrent network比较流行。 : Deep learning在金融界的应用我了解不多,不过很多用recurrent network,尤其是 : LSTM做股票预测的,一搜有好多论文呢,github有很多source code可以参考。 : 高频的散户没搞头,要拼硬件,中长期的可以搞搞作参考,估计效果比看KDJ/MACD也好 : 不了多少。
|
|