Stanford University is giving free online classes. Currently there is one about halfway through (5 of the 10 weeks) but they have another one January next year. Classes presented by Prof Andrew Ng - check out the site at http://www.ml-class.org.
Enjoy
Johan
Ng is mostly known for his distinguished AI paper on Parsing Natural Scenes.
magft wrote:
I'll probably have a go at the 5th NNet over the weekend, so what do you want me to try?
<snip>
Mike
I would start with something very very simple like the close 4 bars ago minus the current bars close (c4-c0) or look for higher highs and lower lows, something like (H2 < H1 and H0 < H1) for a peak or (L2 > L1 and L0 > L1) for a low.
I am an interested follower in this thread - although years ago I supervised a student in fuzzy logic my maths and ai skills are not good enough to make a practical contribution. What I did learn was that the key to using artifical intelligence was identifying the problem that was complex for the human mind but simple for ai.
I wonder if the best use of NN is not so much in identifying the trade but in how to use probability, RR, and money management. Trading a basket of currencies, using hedging, managing DD are fairly complex for the human mind but could be a doddle for artificial intelligence. This could transform strategies like the pitbull's by providing superb management. Opening the trades can be done by the human - I think NN might come up with a really innovative management approach.
Since adding the last post to this topic, I have been thinking more and more about how much I would love to have an intelligent MPTM. I have been developing a trading system which sets up pending orders at the beginning of the week and manages open positions with MPTM. This brings home it is not so much the entry that is the key but the management of the positions.
Steve's MPTM ea is a super addition to the trader's armoury. But could NN turn it into a stealth fighter?
What decisions? - initial SL/TP, changes to SL/TP, whether to hedge and size of hedge, when to close
What variables? - historic range and volatility, trend momentum and strength; balance and correlation of trading pairs.
Please tell me if this is possible or whether I am barking up the wrong tree!
Goog morning Sirs
I started experimenting with NN since one year ago.
Now i'm looking for the "fann4mt_simple.mqh" library but it isn't even available at the given link: http://fann4mt.thetradingtheory.com/
Please, can anyone help me or telling me where to find it?
Thank you in advance
fvtrade
I have read through this entire thread, and I have built an Empty4 indicator using FANN before.
This was after being inspired by Asirikuy's video on Sunqu EA, which is a neural net EA, early
this year.
Although, I do not have a PhD, but I do have both a Master in Engineering and a Master in
Accounting. When I started in Neural Net, I researched several paths and it seemed that
Support Vector Machine (SVM) does better than NN (Torgo's book sorry the title is a bit hazy).
Also, there are other decision-like structured models, e.g. randomForest (used in Kaggle
Competition), that replaces some of the human discretionary function in trading (like NN).
If you had read my last post in Cointegration thread, the R language is quite powerful. Only
recently, an article stated that R data scientists outnumbered all other scientiists using different
statistical languages. Let's cut to the point.
There is a package in R called nnet, that is free and available to use. As the mt4r.dll is already
available as used in the Cointegration EA, there is only a matter of coding several Empty4 functions to
access the nnet package via the mt4r.dll.
But why should we limit our imagination to only nnet, there are also packages such as randomForest, quantmod, PerformanceAnalytics, etc. This will bring the EAs to greater heights. I'll stop rambling
for now.