Strategies based on Neural Networks (or SVMs) - any experience with R ?
I think that Gentil should be aware of the /No Free Lunch Theorem/ (Duda et al., 2001, Wolpert and Macready, 1997). There are no context-independent or usage-independent reasons to favor one machine learning algorithm over another. If one performs better than another, it is owing to its better fit to the particular problem, not its general superiority. If you wish to use these techniques try lots of them: certainly neural networks and support vector machines, but also try some of the ensemble techniques such as bagging, boosting and random forest. You can even try the statisticians favorite, logistic regression. They are all available in R. Stephen Choularton Ph.D., FIoD
On 23/08/2011 12:14 AM, Brian G. Peterson wrote:
On Mon, 2011-08-22 at 10:11 +0200, Gentil Homme wrote:
I just send out this post in order to share within r-sig-finance any possible experience, advice, ... about NNs or SVMs with R.
It seems that you're asking us to share with you, and not sharing much yourself in return. Perhaps you could answer your own questions in this thread with the things you are trying? SVM's have been discussed on this list many times, please search the list archives. This blog has covered this topic: http://www.aphysicistinwallstreet.com/ Also, there are a few books on machine learning that use R.
Several good records have been published in the litterature using these techniques for financial trading strategies.
Which ones? References?
There are also commercial packages (expensive !) which seem to have achieved good results.
Which packages? References again? Note that neural network strategies are very likely to create look-ahead bias as you develop and test them. You try something, fail, and try again on the same data. Unless you are very careful to reserve a 'pure' set of instruments and dates that you won't *ever* touch until you think you have a 'good' machine learning system, you're at pretty serious risk of introducing your look-ahead knowledge into the system. While this is true to one degree or another in any quantitative strategy development, I think it is a particular risk in self-adaptive machine learning methodologies.
So I feel it could be nice to share within this group about the following subjects : - experience using the R packages - data pre-processing before feeding the NNs (technical indicators, wavelets, EMDs, ....) - which type of NNs are suitable - how to build and train them - etc ... Thanks to all for sharing within the R community
Now, your turn. Bring the community up to date with your research so
far.
Regards,
- Brian