Probabilistic neural networks, packages and algorithms for MT4 - page 12
You are missing trading opportunities:
- Free trading apps
- Over 8,000 signals for copying
- Economic news for exploring financial markets
Registration
Log in
You agree to website policy and terms of use
If you do not have an account, please register
Gentlemen!
So what shall we input into the neural network? What error functional shall we choose?
Judging by the content, not many people are interested. Many people think it's about the software....
I suggest you start with the slope of the regression line with different periods. And you can start with different TFs. :)
Functionality of an error - maximum profit.
Linear regression angle to the input, in my opinion, is very interesting!
It's easy to calculate the angle of the ruler (take two points, arctangent and go). But it will be the angle for the given TF. It turns out that for each TF it will be a different coefficient that defines the vertical scale. How do you solve this problem?
Gentlemen!
So what shall we input into the neural network? What error functional shall we choose?
Judging by the content, not many people are interested. Many people think it's about the software....
I suggest you start with the slope of the regression line with different periods. And you can start with different TFs. :)
Functionality of an error - maximum profit.
Linear regression angle to the input, in my opinion, is very interesting!
It's easy to calculate the angle of the ruler (take two points, arctangent and go). But it will be the angle for the given TF. It turns out that for each TF it will be a different coefficient that defines the vertical scale. How do you solve this problem?
It is not difficult to introduce coefficients for each TF. It is also possible to do without coefficients, just scale all values to a given range and send to the input of the NS.
Entering a coefficient for each TF is not difficult. It is also possible to do without coefficients, just scale all values to a given range and feed it to the NS input.
P.S. I went to your forum to register.
Enter the coefficient for each TF is not difficult. You can also do without coefficients, just scale all values to a given range and feed it to the NS input.
I've got a good feeling when i'm trying to open my trading account.
i wouldn't trade looking at one TF... trading in one TF is like blindly crossing the mkad.
on scaling
here's one idea for a medium
take
m1 m5 m15 m30 this is for entering H1 H4 D1 as dominating trend
on m1 m5 m15 m30 you need to catch the full disclosure of the fan at once for 4 timeframes
i.e. MA1 M3 M5 M8 M13 M21 M34 M89 should open a fan of averages at once or start to open it!
by the way the bettor has very similar points!
but in NEUROSET we need to feed something like 0 or 1 for each mean in each timeframe
I propose as an option to take the distance between the averages to bring to 1 if the nearest more difficult below the nearest light
this will be a UP trend on these two averages
when all of the averages on m1 m5 m15 m30 show 1 this is the UP top - then the analysis of the higher TF
i.e. we always start looking for an entry on M1 and then go up to the higher TF
an example of how to scale the distance between the averages
for each average, for each bar an array
..
AdE = 10000;
mas[0][1][ off+ _i ] = iMA(),PERIOD_M1, 5, 0 , MODE_EMA, PRICE_CLOSE, off+ _i );
tmp = mas[0][1][ off+ _i ]-mas[0][2][ off+ _i ]; // between 5 and 8
tmp = MathCeil(AdE*tmp)/AdE; // adaptive simplification
if(tmp>1) tmp=1; if(tmp<-1) tmp=-1;
NN[1][1][_i+8] = tmp; // // // put -1 or 1 in the grid range -1 ... 1
mas[0][0] [ off+ _i ]= iMA(),PERIOD_M1, 3, 0 , MODE_EMA, PRICE_CLOSE, off+ _i );
tmp = mas[0][0] [ off+ _i ]- mas[0][1][ off+ _i ]; // scale between 5 and 3 mas
tmp = (tmp) / Point;
tmp = MathCeil(AdE*tmp)/AdE; // adaptive simplification
if(tmp>1) tmp=1; if(tmp<-1) tmp=-1;
NN[1][1][_i] = tmp; // -1 or 1
I hope you understand what is a full fan opening and the point of beginning of fan opening
PNNNs store almost all data internally - fast learning - but use a lot of memory and are slow
let's say 4 timeframes average 1 3 5 8 13 21 34 55 89 and let's say 5 bars
5 * 9 = 45 neurons for the timeframe according to the given set of averages
45 * 4 = 180 neurons for all TFs
you can try to distribute neurons into M1 M5 M15 M30 layers that will be 4 layers
I would add DIVERGENCY signals to the layer closest to the output
Entering a coefficient for each TF is not difficult. It is also possible to do without coefficients, just scale all values to a given range and feed them to the input of the NS.
P.S. I went to your forum to register.
I wouldn't trade LOOKING at one TF... trading on one TF is like blindly crossing the MKAD
to scale
here is one idea for averages
take
m1 m5 m15 m30 this is for entering H1 H4 D1 as the dominant trend
on m1 m5 m15 m30 it is necessary to catch the full opening of the fan on 4 timeframes at once
Well, this is a separate very profound topic. There are developments in this direction. My soapbox is in the profile.
As for the scaling, I meant a little different.
Here I have sketched an indicator to demonstrate my question that has been bothering me for a long time. It draws a linear regression line. Suppose we want to measure its slope angle, but what scale should we select vertically (by price)?Even on one TF the vertical compression is possible:
In the indicator I have introduced the k-factor for visual adjustment to the necessary chart. In fact the angle value itself has no meaning, what is important is its change. But I would like
to be a value (not necessarily angle in degrees) having the same scale of variation for any TF.
I think that mathematics solves this problem one way or another.
I cannot say anything about neural networks. I've never designed them myself (especially in C) but I would like to, but I have no time.
P.S. I like your Expert Advisor on divergences. I wish you a good finish at the championship.
can't find NeuroDimension NeuroSolution 5.06 Developer, can anyone... ...at least give me a hint. andrew.opeyda(dog)gmail.com
I have one:
E views
Poly Analyst 46
Evolver 4.06
2 Paramon
can't find NeuroDimension NeuroSolution 5.06 Developer, can... anyone... ...at least give me a hint. andrew.opeyda(dog)gmail.com
I have one:
E views
Poly Analyst 46
Evolver 4.06
Best to get it from the developer's website, but you'll need a little registration.
How to prepare data for neuron input!
Suppose there is a neuron with three inputs, each input has a W scale
neuron needs to output a value
1st choice neuron receives some range of already transformed data let us say { -1.0 -0.9 -0.8 -0.7 ... 0 . 0.1 0.2 ... 0.7 0.8 0.9 1. 0} for each input
there are only two 0 : 1 output values
Option 2, the neuron receives some range of data already transformed let us say { -10.0 ... 0 ... 10.0 } on each input
the output is also the same range of values but including weights
option 3 receives { 1 0 0 0 } on each input in the output depending on the weights { 0 1 }
how do you even prepare the correct conversion data.... for a neuron... it can't just be 1 and 0 ... there has to be a range ?
I'm talking about the incoming layer! Each layer shrinks the data more and more
The idea is to have 6 states in the output of the network and not just 1 and 0
at the output we have 6 states
1 1-sell
2 1-close sale
3 1 buy
4 1-close buy
5 1-hold buy the rise trend
6 1Hold sell - decreasing tendency
maybe i am wrong
often if the function is a hyperbolic tangent, the inputs are normalized to -1...1 or 0...1
but who in neurosolutions compiled the dll?