Better NN EA development - page 3

Wayne Ovenstone
992
Wayne Ovenstone  

References?

Hi Mark

That made me smile, thank you.

References, LOL.

Why would anyone need references on you, if they do then they do not deserve your abilities.

Keep coding bud.

Wayne

Orest
21
Orest  

Problem is solved. Enjoi the solution....

Hey guys,

you don't need to re-invent the wheel...

Please read how to create Metatrades based EA using Cortex neural networks software. It's basically step-by-step instruction which is pretty easy to follow. You even don't need to use cortex, but it gives you general idea how this kind of EA should be built. I found there very similar ideas Olexandr mentioned in his contest interview (for example C++ optimization before even using MQL)... The only problem with this approach is - again as Olexandr mentioned he teaches his EA as he trades... it seems not clear how to do that with pre-defined optimized coefficients... Anyway, read and I'm sure you will get much clearer idea.

Neural Networks and Stock / Forex Trading

tiger_wong
150
tiger_wong  
Orest:

... it seems not clear how to do that with pre-defined optimized coefficients... Anyway, read and I'm sure you will get much clearer idea.

Neural Networks and Stock / Forex Trading

I think automated optimization could solve the coefficients... what do you think, guys?

barnix
1310
barnix  

Support Vector Regression and

Neural Networks in Time Series Prediction

Files:
dmi5594.pdf  202 kb
nn6.jpg  35 kb
[Deleted]  

Ive been doing this for a week

I ve been developing a NN based EA for about a week now. Mine uses an external dll, it is running in the strategy tester optimizator as I write this. I strongly suggest you keep it simple, because very complex models may work on test data but fail in live trading.

"predicting MA"s makes no sense; predicting price is more like it... but if you read the literature on the issue (You could start with edward gately's "neural networks for financial forecasting") you'll soon understand that predicting change is less error prone.

If i am not wrong, Mr. Better said that he uses Moving Averages as inputs,not that he predicts them.

There are a number of tools that will actually train and write the code for you: statistica's neural network module, neuralworks, tiberius... im just afraid that their predictive results are a bit overrated

someone mentioned there's another thread on neural nets; where is it? I have 4:43 more hours to wait until this optimization run finishes...

bolt
39
bolt  

There are too many limitations in the MT4 framework for intensive NN routines. The training is best handled externally so simply collect an array of the last 1000 bars together with the indicators of your choosing perhaps 3 or 4 is good and certainly needs more than one. Dump the array together with indicators values to cvs or mql and your tuning there instead. Auto split 80% learn to 20% test for validation. There are dozens of excellent c++ libraries which can tune and model this array faster then mt4 and without limitations. Finally only the wights file needs bringing back to Mt4.

However NN on its own is no good because with a GA input selector you have no idea what produces the best signals. As already in this thread you are discussing what MA's should be used if any at all?? GA selector will find the best inputs. That said in my experience GP is MUCH better then NN for financial time series and faster too. NN suffers from hill climb and easily bound to local minima providing a too easily curve fitted solution. GP never suffer this problem either they find a global solution or not. So in my opinion you should look closer to a genetic solution to input selection and global search and the resulting weights file is often much leaner and more efficient for MT4 to use after. I suspect that BEtters NN system has simply caught a sweet spot the last few weeks but you see prior he was down in the noise with all the other EA's.

DeVinci
285
DeVinci  
bolt:
There are too many limitations in the MT4 framework for intensive NN routines. The training is best handled externally so simply collect an array of the last 1000 bars together with the indicators of your choosing perhaps 3 or 4 is good and certainly needs more than one. Dump the array together with indicators values to cvs or mql and your tuning there instead. Auto split 80% learn to 20% test for validation. There are dozens of excellent c++ libraries which can tune and model this array faster then mt4 and without limitations. Finally only the wights file needs bringing back to Mt4. However NN on its own is no good because with a GA input selector you have no idea what produces the best signals. As already in this thread you are discussing what MA's should be used if any at all?? GA selector will find the best inputs. That said in my experience GP is MUCH better then NN for financial time series and faster too. NN suffers from hill climb and easily bound to local minima providing a too easily curve fitted solution. GP never suffer this problem either they find a global solution or not. So in my opinion you should look closer to a genetic solution to input selection and global search and the resulting weights file is often much leaner and more efficient for MT4 to use after. I suspect that BEtters NN system has simply caught a sweet spot the last few weeks but you see prior he was down in the noise with all the other EA's.

Hi everyone and BOLT

Well, maybe it would be interesting for somebody to explore the following site ( lots of FREE Genetic Apps.... ) :

Sales

Download OK

Software

Genetic Pattern Finder

Bye

DV

jdpnz
47
jdpnz  
alpha24seven:
It has been my experience that using NNs to predict a MA is very difficult. There is simply too much variability (noise) to deliver a profitable model. This is just my experience and I'm obviously wrong if this NN model is going to win the EA contest.

In my experience, it is possible to "predict" a profitable sma (but not ema/hull or any simular) using the jurikres object (or your own) to built a proper data set.

This is because a sma has a fixed lag - 50% - and thus also a "predictable" velocity curve while most other ma's do not.

However - the output signal becomes more noisy (difficult to control) from about 25% lag towards 0% (real time)

I have never been able to actually create a real time predicted ma yet.

However - in actual live trading, I use some interesting combination of a cycle measurement (to determine cycle periods - something reminding of the Ehlers dominant cycle) and this cycle period I then use in combination with some wavelet formation.

Used properly, the wavelet provides me with a 12% lag (constant and very reliable) which then I feed to some NN. Using this, I create a MA with 5-7% lag only (better than a ZMA even) with minimum noise - and no overshoot like a ZMA.

This I then use as a basic building block for building indicators - in some indicators this is indeed very useful.

jdpnz
47
jdpnz  
hunter113:
Hi,

What we know about Better system: (at least what I realized:)

-It is trading 2-3 times a day, which is something I really like.

-Using static stop loss of 60, and taking 30-60, sometimes more

pips profit.

-It is using PNN, input some moving avarages, (what time frame,

what length is question)

-It is built up from 3 subsystem (maybe this is the 3 inputs fxspeedster

referring to)

So for me it seems that it is predicting the most probable next 30-60

movement.

Hunter

This is interesting and would suggest to me that this particular NN has been trained much like Barnix suggests - with perhaps a 50 pip Target area and a "weight" (trend strength) depending on the relation between wrong/correct direction before the next 50 pips are hit.

In this, actual input data (MA periods etc.) would also become less important - because the system would train towards the above relationship. As such it may be well worth exploring this idea a little bit more.

jdpnz
47
jdpnz  
bolt:
I suspect that BEtters NN system has simply caught a sweet spot the last few weeks but you see prior he was down in the noise with all the other EA's.

If this is the case it would suggest that fixed (period) ma's/indicators have been used as input data - this could be resolved by implementing some cycle measurement subsystem underneath the NN development.