Machine learning in trading: theory, models, practice and algo-trading - page 822

 
Alexander_K2:
In a week I will post BPs already reduced to Erlang's flow of the 1st order. But I don't think I'll do 2 and 3 - I don't need them.

better in the form of code (pseudocode) for the indicator, I pray! then you can do a full training on the different characters and look at

 
Alexander_K2:

YES! I'm thirsty!

Right now I need to tap these rows with neuron, and then only give them to the coders.

And I don't have the time right now. I'm telling you - our "partners" from Germany are forcing me to do some project, while I've abandoned the Grail. There it is - in the corner...

You don't need it there.

The main thing is not to fill your head with neuronics in your case

In fact, you got a corrupted exponent, and not only, a statistical graph of the vectors of increments.

Go back to vectors, you don't need any threads.

Well, what is the direction - the sum of vectors (of course, you should think about which ones)

That's all

Good luck!

 

Recommended for study (from fxsaber, he will not advise bad)

To the question of stationarity of influence on the target and the "right" chips

https://www.mql5.com/ru/forum/232030/page2#comment_7026700

Скрипты: ThirdPartyTicks
Скрипты: ThirdPartyTicks
  • 2018.04.06
  • www.mql5.com
ThirdPartyTicks: Автор: fxsaber...
 
Renat Akhtyamov:

You don't need it there.

The main thing is not to fill your head with neuronics.

In fact, you got a corrupted exponent and not only, a static graph of vectors of increments.

Go back to the vectors, no need for any flows.

No. It's a dunce or a dunce. Only unrestrained sifting through BP can bring cash - I stand by my opinion.

 
Aliosha:

returns with windows e~2 -> 1,2,4,8,16,32, 64.... 10 pieces, some take more meaningful 1,2,5,10,20,30,60,120,300,.... But it is not essential, at the expense of regression and other filters, it is also considered that it does not affect markedly, it's a matter of taste, the main thing that the data were synchronized and did not wander in relation to each other when there are a lot of series (and there always are), otherwise everything breaks, so you have to write them yourself, when there are a lot of sources, for example returns from the past {-1,-2,-4,-8,-16,-32,-64,-128,-256,-512} and target ruturns from the future {+1,+2,+4} as a "stylized" example, more volatility changes there, there are also changes in volatility, delta of cup, OI, etc. BUT the MAIN thing is to use targets with NOT TRANSFERABLE DATA, otherwise it's a fake, which can easily be improvised by any peeping turkey like ZZ

If serving such returns, but should they be sifted by the effect on the target? As in the articles the predictors obtained from the indicators were sifted out.
And, for example, leave 2,16,64 on prices, 1,4,32 on volumes, etc.? (By the way, synchronization will be out of sync here, since different bars may appear).
Tried without sifting - the result is about 50%.
I haven't tried it with sifting... If anyone has tried it, are the results better?
 
Aliosha:

returns with an exponential window, for example, and anything else, RSI, makdak, etc. "WITHOUT retraining" of course, on the test, but since ZZ retraining is inherent in the way he looks in the chips, even on random wandering you can easily get 90% on the test.

Either you made a model without retraining, proving this fact in the test, or learned about retraining on a depot drain. Personally, I find out about retraining on the test. The method is simple, but it is carefully avoided by the local coryphaei.

ZZ does not look anywhere by definition: the last link does not exist at all, and the penultimate one changes from time to time. ZZ is a visual representation of the trends with a formal definition of a trend: the number of pips from the last turn.

But ZZ has another problem, at least for me: I was unable to fit it with predictors, i.e. I can fit the model, but I cannot create a REFORMED model. And the reason is not that ZZ is looking somewhere, but that I could not pick up predictors that would have sufficient predictive power, i.e. my predictors for ZZ are noise, on which and manages to fit the model with very good indicators (error less than 10% easily), but if you run this model outside the training-testing file, you get random error, which is an indicator of retraining.

 
SanSanych Fomenko:

ZZ looks nowhere by definition: the last link does not exist at all, and the penultimate one changes from time to time. ZZ is a visual representation of trends with a formal definition of the trend: the number of pips from the last reversal.

ZZ does not look forward when working on a real account or in the tester on the 0-m bar, but if you look at it on the history for example on 10 000m bar, then it certainly does. The matrix for learning is taken from the history.
That it looks into the future is a good thing for the target as we want to predict the future.

On the other hand, Alexey says that it cannot be used as a target because it is looking into the past as well. So what? We only feed the data from the past to the input of the NS.
 
Aliosha:

Exactly! You can't look into the past for targeting, just like chips into the future, if you gentlemen don't understand WHY, do an experiment with SB,

To understand WHY - you need to have a logical explanation, your words do not show it. Could not explain, without referring to the experiments?
Alyosha:

On SB a correct ML should not give stat significant deviations from 50% (on the test) with ZZ targeting they will be because the classifier is adjusted to the past contained in the target, hence the space accuracy, but not the fancy Sharpe

To fit to the past, it's enough to skip any input to the output without any transformations.

 
elibrarius:
To understand WHY - it is necessary to have a logical explanation, from your words it is not visible. Could you explain, without referring to the experiments?

To adjust to the past - it is enough to let any input to the output without any transformations.

I don't understand what you've written at all - it's nonsense.

 
Aliosha:

Exactly! You can't look to the past for targeting, just like chips to the future, if you gentlemen don't understand WHY, conduct an experiment with SB, I wrote above and have written before me, I repeat:

conduct an experiment with SB

conduct an experiment with the SB

do the experiment with the SB


on SB the correct ML should not give stat significant deviations from 50% (on the test) with ZZ targeting they will because the classifier is adjusted to the past contained in the target, hence the space accuracy, but a lame Sharpe

Don't give the assignments to others, get the results here.