Welcome to NexusFi: the best trading community on the planet, with over 150,000 members Sign Up Now for Free
Genuine reviews from real traders, not fake reviews from stealth vendors
Quality education from leading professional traders
We are a friendly, helpful, and positive community
We do not tolerate rude behavior, trolling, or vendors advertising in posts
We are here to help, just let us know what you need
You'll need to register in order to view the content of the threads and start contributing to our community. It's free for basic access, or support us by becoming an Elite Member -- see if you qualify for a discount below.
-- Big Mike, Site Administrator
(If you already have an account, login at the top of the page)
Has anyone thought about trying NN's by using market replay to simulate the "perform" phase (after training)? I have a load of market replay data. From my experience, forex and other liquid markets are the best for NNs. I can upload an update of my market replay data. Just remember that forex volume is rounded. So instead of 21 million per minute, it will just display 21. This can be fixed by creating a new volume indicator that has a factor programmed into it (such as 1000000).
Have done so in the past (when I was more enthusiastic about NN techniques than I am now ), agree market replay has the advantage over real time forward testing that time can be compressed, and using a trading platform to evaluate overall performance of a strategy that depends on NN generated signals is probably more efficient than reinventing the wheel in MS Studio.
However, to the extent strategy performance can be estimated in code (or in Excel, say) from signals generated by an NN under test and perhaps 99% of the effort required to develop an NN based strategy is determining net architecture and feature vectors that generate usable signals, during development I find it more efficient and less cumbersome to import all available data into the NN app (which is designed to test as well as train), use part of data for training and the rest of it for testing.
In other words, IMO market replay may have a role in the end game but it is minor.
All, I've had no success trying to apply NNs to shorter time frames & other instruments than have been described in the literature. but in my opinion this is because of lack of trying rather than disbelief in the possibility it can be done.
I strongly suspect my problem so far is focus on code development and NN topology (NN architecture) and not focusing on feature vector development (input components that have genuine correlation with outputs), which is essentially to reiterate if you don't know how to trade your chosen instrument in your time frame you don't know how to teach an NN to do it.
Which is why the largest part of my energy these days is expended as follows, rather than messing with NNs:
1. translating a proven trading method to algorithms a bot can use
2. extending the method to as many instruments as possible.
I have been focused on different Machine learning Systems as well as researching prediction of Time Series. I have a prototype of SVM (Support Vector Machines) which illustrates to me the research is correct on time series. No wheres near perfect but pretty good results on the first short training example. I see the Wave59 product also showing hope in this area on day trading, but feel it is subject to noise during the day and doesn't perform well with the higher noise to signal ratio of Minute bars.
Some of my early work was based upon dumping data into a black box and expecting a result. This did not work well for me either. I think the Artificial Bee Colony had limited success, this was my earliest work. The success was captured in a statement from an expert in AI and was a little bit along the lines of "the system works in spite of the programmer". Mainly because it is so complicated (multi dimensions) it tries to do what you want, but maybe not reliably.
Unfortunately, I don't have an enormous trading background. I know very little about market conditions and trading system and have done some much less in actual trading. Most of my trading has be simulated through Bots or a little "ad hoc" discretionary trading. This is what drives my Machine Learning interest. If I can determine after it happens when a trade was good, hopefully I can teach a computer to recognize that and do it in the future. And then one step further, help that computer "generalize" that ability to recognize other similar situations.
My research into time series has really helped, there are many research papers out there on forecasting/predicting time series that clearly illustrated you need to approach the problem with the right hypothesis or it simply isn't going to work.
I am hoping to bring more hope into Machine Learning and attract more thought in this area. I am sure I will check back with you as I move forward (hopefully with good results).