Neural network - page 9

New comment
 
Vanek_MIL писал(а) >> And if one compares (under the conditions given above, of course) the original signal and the derivative, is the derivative the choice?

It all depends on what the original signal is, what its derivative is and what we want to get as a result. But of course, the less transformations the better, because each transformation introduces additional distortions to the original signal, which of course can have a negative impact on the final result.

Vanek_MIL wrote(a) >> And if one extends the situation to crossing some threshold by setting the offset relative to zero, should the "threshold" signals be amplified in this way...?
What do you mean by "amplify the threshold signals"?
 

LeoV писал(а) >>

What do you mean by "amplify threshold signals"?

I was referring to a derivative signal that peaks when the original signal crosses some threshold - i.e. a generalisation of the above example.

LeoV >>:

It all depends on what the initial signal is, what its derivative is and what we want to get as a result

.

So, it all comes down to the task at hand.

Allow me to bring my way of thinking - how and why I am going to use nets.


  • When training a network using an indicator that produces ready-made signals, I expect that the network will find some combination of inputs (pattern) preceding the occurrence of the signal and predict the signal when the pattern repeats.

How will it "know" that it is the one that precedes it? Well, probably because it will give more profit in a strategy based on these signals. Although not unambiguous here of course - just a guess.

The pattern must (? may) include the static component (values) and dynamic one (momentsums, etc.) to complete the picture.


We have already discussed above learning with the teacher. The indicator will be the teacher, which gives good signals (and after optimization - even better).

But then why not use an "ideal" teacher giving signals on the basis of the same zigzag. Although here the question is - can the teacher be allowed to look into the future without allowing the network to look into the future?

  • It is possible to use training without the teacher relying on the fact that the network will find a state, in which signals given out by it maximize profit. But it is hard to predict anything here - just choose (?) the type, configuration of the network and set of inputs?
  • When working on the selection of inputs the indicators are selected/constructed that more or less reliably reveal some point (approaching the consolidation level, for example, or lower correlation between the pairs), which by themselves are not enough for the signal formation. Is the network able to process this whole set of data and draw a conclusion? In other words, what qualities must these signals have for the network to work effectively?


Regards to those gathered.

 
njel >> :

The idea of the TC must be present.

To complete the picture: the idea is to use the network to receive entry signals (or their confirmation), and then to accompany the trade and exit "independently" - either by stop-loss or by receiving a reverse signal.

 

There's talk about learning with and without a teacher

And a question arose - to users Neuroshell (perhaps in other programs is similar - I do not know just) - regarding the addon Neural Indicators.

I will make some suggestions (if wrong - correct).

If we take usual neural networks based on Turboprop (either Predict or ATR2), or probabilistic ones - then training is going with a teacher. As a teacher we take some indicator - or standard Neural Outputs like Optimal%Change, or price change, or Bai/Sell flag, or something based on the same zigzag (if that is what you mean) etc. So, there is an assumption that teacher's signals should be somehow coordinated with input data. Otherwise there may be situations, such as input gradually increases, then decreases, but the output remains constant. Or input is constant - output increases rapidly at the beginning and then decreases. There are many variants and everything is much worse when there are a lot of inputs. And things like that are likely to lead to network stagnation - because the same inputs lead to different outputs, or vice versa.

So the upshot here might be that picking the RIGHT teacher is very difficult, if we make a mistake we risk ruining the network even with good inputs.

Possible solution is to use nets from the neural Indicators addon - they are trained without a teacher and are adjusted to target strategy functions

Question - does this addon have a distinct advantage over other neural networks?

 

I don't understand why we need to take derivative quotes when we can use a neural network to predict the possible direction of movement by inputting a series of previous values of HIGH CLOSE LOW

like 400 bars deep ;) ? On H1 timeframes and less it is necessary to take into account the OPEN price. This makes 400 X 4 = 1600 HCLO input values for M1: 60 bars ahead will be enough for predicting the direction.) All that remains is to find a suitable analyzer program and a supercomputer.

 
Piboli писал(а) >>

I don't understand why we need to take derivative quotes when we can use a neural network to predict the possible direction of movement by inputting a series of previous values of HIGH CLOSE LOW

like 400 bars deep ;) ? On H1 timeframes and less it is necessary to take into account the OPEN price. This makes 400 X 4 = 1600 HCLO input values for M1: 60 bars ahead will be enough for predicting the direction.) The only thing that remains is to find a suitable analysis program and a supercomputer.

The most important thing is to find the right analyzer program ))))

>> Excuse me, of course, but have you had a good experience with this approach?

 
GrooovE >> :

The most important thing is to find the right analyser programme )))

Excuse me, of course, but what - has there been positive experience in such an approach?

Well on H4 for 4-5 bars ahead perc for 90 hundred reliably yes for 2-3 weeks without retraining...

I can't find a normal program, besides Forex I have to study www.wasm.ru ;)

I have to study Forex and I don't know how to do it.


 
Vanek_MIL писал(а) >>

To be honest - I don't get it...)) On the subject of the perfect teacher - it's not at all certain that the perfect one is needed......

GrooovE wrote(a) >> Question - does this addon have a distinct advantage over the other neuroshell nets?

One advantage - no teacher.....

Piboli wrote >>

I do not understand, why take derivatives of quotes, when you can use neuronet to predict the possible direction of movement, by giving as input a series of previous values HIGH CLOSE LOW

like 400 bars deep ;) ? On H1 timeframes and less it is necessary to take into account the OPEN price. This makes 400 X 4 = 1600 HCLO input values for M1: 60 bars ahead will be enough for predicting the direction.) All that remains is to find a suitable analyzer program and a supercomputer.

The problem is one, when the price goes beyond the range that was on the history of 400 bars, the neural network will not know what to do and will give a signal in the direction of this range regardless of how the price moves outside it......

 
LeoV >> :

Honestly - I don't get it...

As they say, a properly asked question contains most of the answer. Therefore, even from such an answer certain conclusions can be drawn...)


Forgive me for my annoyance and let me introduce one more reasoning question:


Can a neural network be used to determine the current market phase?

Let me clarify what is meant.

By market phase in this case we mean the presence of a trend or its absence (probably it is a flat).

Ok, let's leave the teacher aside. Let us take the net without training.

So inputs must be selected in such a way that clusters (forgive me for my incompetence), at best two of them should appear in the input parameter space:

One for a trend, the other for a flat... (or maybe only one cluster is a trend, and everything else is not a trend?).


If that's the case:

How, in what form does the output signal show belonging to a particular cluster?

Is it realistic to visualise these clusters to navigate already at the input selection stage?

When designing such a network, is it possible to control the process methodically or one has to rely on chance - will it work or will it not work (of course, this is a naive question)))?


Respectfully.

// went to learn the maths

 

On the subject of backprop, I would like to add that using the algorithm's standard error is wrong.

1234567891011121314
New comment