Neuroshell has some nice features, but the included neural network functionality is very basic, nothing that would give you an edge in and of itself. An NN that is actually useful is so far removed from the basic networks in neuroshell that it would be just as easy to implement in Ninja, probably easier actually considering that ninja is more extensible.
In machine learning, the learner is only as good as the teacher. In other words, for it to be useful you have to already have an 'edge'. Machine learning techniques can help you uncover details about relationships between things or patterns in data but in order to find anything useful you have to specify what to look for and where to look..
It is ultimately the same as optimization, you can take a good set of rules and make them better but unless the problem at hand is very simple you are not going to extract good rules out of thin air or take bad rules and make them good.
Broker/Data: Tradestation/Tradestation, NinjaTrader, FXCM and Tallinex
Favorite Futures: ES, CL, EUR/USD, TF
Posts: 171 since Aug 2009
Thanks: 99 given,
I think the answer lies in a way to interpret support and resistance in terms of Demand and Supply respectively. Sam Seiden was one of my teachers from Online Trading Academy and if you search the internet, he has lots of articles and webinars on this subject. although, he believes this cannot be programmed, I think otherwise since it is very objective.
Anyway, if the market lets say sold off and bounced back from a point very quickly in terms of price and time, the origin is considered a good demand point. So, the 1st time down in this area, you would want your strategy to first be able to notice this, and then dont take any short trades in demand zones. The 1st time the market hits the zone is the right time to take the trade because there is more demand MASS there and each time price revisits, the mass gets less and less.
So, I think if you could teach a strategy how to quantify supply and demand zones, filter them through trend of different time frames, and give these zones strength levels, things would be pretty cool.
For example: On 10/10/08, the marekt bottomed in the crash. It rallied for a few days then revisited 825 again on the ES on 10/27/08. This was previous demand. It was demand because the last time it hit this price on 10/10/08 it rallied to 1050 before falling back down. It is not guaranteed, but there is a good chance because of that POP from 825 on 10/10, it would POP again. However, because we were in a downtrend, you have to discount beating 1050
So, perhaps an AI would notice a swing low and have a grading scale based on current volatility per time and price filters to grade the quality demand or supply. It can do this on the daily, weekly, monthly as well as 30min right now to range and tick charts or volume charts, etc.. So, if the demand of monthly, weekly, daily, and hourly line up, with Timing arrows, this is a great time to go long. From our emotional point of view, it might not feel this way and we might wait until it "feels" better, but risk might be too high.
So, that would be interesting to build.
Also, Prime-Line trading has this Matrix thing that is an automated thing that seems to find confluence areas. Basically, I think if an AI can find confluence areas, then timing arrows and strategies would probably work very well on a consistent basis----especially when you employ 2 or 3 contracts and shed the 1st 2 contracts with targets and then let the third one ride.
I have neural network code for Laguerre-type nodes, but we would have to adapt it to C# and figure out how to let Ninjatrader modify the values. I am not entirely sure how we would get the data from the live feed into the network for calculations, and we would have to work out a way to keep from reloading the live data for each new neural network calculation (only add the new bar instead of all of them)...
I've started to use Ninja just because it is .NET based. Why I need this - because I love C# and I use it more then 5 years. I need it for trading, because I'm developing neural networks based system. If anyone need libraries you can find in codeproject.com. i can't post links as I am new.
These have everything you need to implement NN.
I think that NN can and are useful in trading. I'm so sure because my fried made master degree graduate work in NN subject and he tried to forecast stock market. And it wroked. The theoretical profit was about 0.5 percent per day. Without leverage! This sounds too good. But I'm developing it and I hope to test it soon.
You can get live data from Metatrader, that's for sure. I think you can get it from Ninja too. I will find out that. As you have trained neural network you don't have to train it again after a new values arrive. You just put this data into NN, values are multiplied with NN weights and you get output data - forecast. It's simple. The hardest part is to choose layout of NN and input data type. Tick data would be too fast. I'm thinking about M30 open, close, high, low, volume. So there would be 5 input values.
I have been trying time and price with 5-7-5 and it seems to find highs and lows quickly, the part I do not fully understand is the impulse factor of the Laguerre, it appears to basically be trailing off from five levels, so I am going to try to force the impulses past the halfway point of the impulse to be on the same side and see what happens to the results. I do not want to unintentionally recreate a guppy system just to make the impulse ranges look pretty, plus I think the Laguerre ends up being too strict on tick data - - it can segment just before or after a large bar, and that is what I am trying to avoid.
What do you think of Streambase - - I have not tried their demo, but would like to implement two categories of rules to create a warning system of sorts - - one that accounts for market hours and holidays, etc. and one that parses the news events (like they go into detail about in the video).
What is your opinion, and / or do you recommend a better method?
The following user says Thank You to turbolover for this post:
Totally agree. From what I've looked into this seems to be a waste of time without a very thorough understanding of AI/ML...
I read a story of an early AI experiment to find soviet missiles from spy photos...Researchers couldn't figure out why the algorithm was working so poorly until they realized that the training photos were all taken on sunny days so what the algorithm really ended up doing was identifying sunny vs cloudy days.
With so much market data at hand you have this problem X1000 and in the end will probly just end up building a spurious correlation engine.