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)
NinjaTrader with Neural Networks using Encog framework
For all NT users (capable of programming on the NinjaScript level) who would like to dwelve into the world of neural networks – now you have a chance for contribution.
Jeff Heaton has just released example project integrating NT7 and Encog neural networks framework.
For sample project and article go to:
Hopefully futures.io (formerly BMT) community can join the efforts and help the project evolve into something even greater for the benefit of both NT and Encog.
If you have any good ideas for the project we could work on I suggest you answer below and if there will be enough people interested we could start doing it together. If you have a project in mind that you would rather put in the Elite Circle let me know so I could open similar thread there.
Heaton Research have also provided a number of educational videos on youtube.
I have found Encog great for prototyping ideas as it supports quite a number of different neural network types i.e. feed forward, recurrent. It also allow for the use of genetic algorithms for network training.
I'm currently looking at using it to predict the ROC for the next N days. My next project is to try and use a network to classify significant turning points.
Jeff Publsihed a new article , how to develope a neuro indicator for ninja, now it should be easier to develope new type of indicators, using wavelets or combinations of indicators. In example it reads indicators values from csv file which ninja outputs, and learns,then outputs neural networks what ninja indicator can use.
If somebody is interested to starting developing workbench around enqoc, using enqoq for a new type of indicators and have a background of using AI, send message.
Anyhow it should be easy to develope indicators now, but needs stuff around.
I think enqoc have potential because it can use GPU's and "clouds"
I imported Encog into NT 7 tonight as an experiment, the plan being to use SharpNEAT/HyperNEAT as the engine eventually if anything comes of it. I had to rebuild the Encog archive downloaded from Jeff's site with the DLLs in the top directory of the ZIP file since it wouldn't import as is, but otherwise appears to function as advertised.
I used neural nets for time series analysis years ago and am more or less aware of the limitations and applications, but the idea this time isn't necessarily to try to develop a neural trading system as to get a better grasp of strategy optimization by comparing it to something I'm familiar with (I'm new to auto-trading).
This occurred to me because my strategies are turning into a sort of of "neural" network anyway, with switches (binary parameters) to turn code segments on and off or wire different indicators together (to modify the "topology" of the strategy so to speak), and other parameters to weight indicator output (to make trade decisions--buy or sell & how much--based on some measure of "certainty" or a kind of "fuzzy centroid"). As usual the output of the optimizer is a "best choice" vector of parameters, in this case defining the optimal code & indicator combo and weights of the connections between the components.
The problem I foresee with this approach (optimization in general and neural networks in particular) is how to avoid producing a strategy that is "over optimized" (a strategy that works well with the training set yet fails miserably in real-time simulation, say), or at least how to recognize it, since once trained an indicator-based strategy is only as adaptive as the indicators comprising it.
Small set of variables. (inputs), neural networks means also number of layers and hidden layers, genetic algorithms and evolutionary systems, means, number of locus in chromosomes.
Different "models strategy" for longs and shorts.
Leading indicator, example trin, adv/dec for ES, or any instrument which correlate with x. SP500 index for stocks.
Sometimes removing peaks. and replacing average/median high or low value.
Sometimes adding noise.
Sampling, sometimes learning most recent cases, and veifying historical data.
Random sampling, learning dataset. where is data is randomly distributed over learning set.
So there is recent datapoints and also historical datapoints in learning set. that way you avoid, learning set have only price information when market is going one direction, and testing data is going other direction.
And good Fittness Functions or how to measure which models or strategies are best, very important.
