Gathering a team to develop an IO (decision tree/forest) in relation to trend strategies - page 8

 
Roffild:

The whole thing.

Now it's more like, "Guys, I heard that scaffolding and nets are cool! Let's try them!"

First, study the tools themselves.

Yeah, I suggested that in the machine learning thread, I even made a script to generate pseudo-random but fairly periodic timeframes... Who the fuck guessed that you should first test the mapper on data you can understand how it's generated and after making sure that the mapper can work with this kind of input data - that's when you can "go for the wood

sad all this ))))

 
Aleksey Vyazmikin:

I propose to rally to the challenge of MO in relation to trends, i.e. where the probability of an outcome event must be predicted many bars ahead.

What I propose from my side:

1. a basic strategy for generating targets.

2. Maintaining a joint database of features.

3. Organizing the possibility of optimization/other calculations which require computer power. If necessary, capacity expansion. Possible creation of a common network or optimisation on behalf of.

4. Conducting tests in the strategy tester in order to detect errors in the algorithm's logic.

5. Testing of strategies on demo accounts/cent accounts/other accounts with small lots. If necessary, I will provide a machine for VPS (though I will need help in setting).

6. Holding collective meetings to develop common goals and activities.

7. Providing ideas for building a decision tree.

8. Coordinating the work of the team.


Who is needed:

1. Programmers able to work with MQL5

2. Programmers with skills in programming languages like R and/or Phyton.

3. people with knowledge of MO and/or statistics 4.

4. others who wish to contribute to the project and are willing to offer something - the matter is negotiable.

Sign me up to the group.

 
Roffild:

The whole thing.

Now it's more like, "Guys, I heard that scaffolding and nets are cool! Let's try them!"

First of all, study the tools themselves.

If you have read the thread carefully, you should have noticed that I am not just talking about hypothetical application possibilities, but already applying my method in practice. Therefore, comparisons are not entirely correct. However, I am well aware that my knowledge is not enough, so I want to surround myself with people smarter than myself. To provide an opportunity for cooperation. Acting as a coordinator, rather than a lighthouse of thought.

And, once again, I would like to tell you that the basic toolbox is in need of modernization - it is not efficient to perform operations using only the saw, you must use the scalpel in many cases. Both tools cut, but the result is different...

 
Igor Makanu:

Yeah, I suggested this in the machine learning thread, I even made a script to generate pseudo-random, but fairly periodic timeframes... Who the fuck guessed that you should first test the mapper on data you can understand how it's generated and after making sure that the mapper can work with this kind of input data - that's when you can "go for the wood

sad ))))

Don't be sad about other people's stupidity - get busy and enjoy your achievements!

I don't understand people who only say things are bad but do nothing to change the situation, even when they are openly offered to do so.

 
toxic:

Sign me up for the group.

Great! (Laughs)

Who are you willing to join as?

I know about your background in MoD, but I don't know about programming - what language do you work in?
 

On the subject of reinventing the wheel, I assume this has been thought of before and used, if so please let me know.

So, I'm an advocate of visual analysis - that's how I learn information better, so I thought, what if the tree sheets are presented as a matrix? I.e. we have a sample of N rows and these rows have patterns for rules execution in the sheet, then we can expand sample length by different proportions, for example width will be days of week and height - week, or width - specified percentage of bars per cell, and height - month. The cells will show the presence of a rule (pattern) and the result of the classification. By this representation one can see how the rule is spread in space and estimate its stability in time. Of course, for these purposes we can use not only our eyes but also a mathematical apparatus, but visualization is required in the stages of algorithm development and hypothesis making. I assume that if a rule is spread evenly over the entire sample, rather than concentrated in a single time interval, then such a rule is a more stable pattern than a rule that was heavily concentrated in any range, and we should also take into account how the predictive power of the rule changes over time. Also this method will allow to identify features with rare combinations, and hopefully these combinations can be excluded for genetics or removed at an early stage. Also this method will allow to see the cyclic pattern, if there is one.

And, I just believe that it is necessary to keep consistency of time series when estimating results, while classical approaches (from what I saw in the form of lectures and read in the form of articles) do not see the need for this, as usually independent situations in the sample are used.

And if arrangement in space of patterns may be used (it gives positive effect), then we may teach fitness function to try to build a tree in order to improve results of forecasting and to find more global regularity. But probably here we should not apply the greedy principle, but some other approach - we need to think.

 
Aleksey Vyazmikin:

Glad you are interested in collaborative activities.

How do you organise these groups?

However, as long as there are hypothetically only two people it is difficult to talk about a group, i.e. for now we should wait to go into a closed space, maybe someone else will want to work in this direction.

Here on a site in messages, the truth the group itself it is possible to create only from the terminal from a phone for some reason :) All sorts of automatons and other battered old guard makes no sense to listen, never heard anything useful from them
 
Aleksey Vyazmikin:

And, I just think that it is necessary to keep time series consistency when evaluating results, while classical approaches (from what I have seen in lectures and read in articles) of building a tree do not see the need for it, as usually independent situations in the sample are used as it were.

And if arrangement in space of patterns may be used (it gives positive effect), then we may teach fitness function to try to build a tree in order to improve results of prediction and to find more global regularity. But here it is probably not the greed principle that should be applied, but what

Genetics is a pretty useless thing in TC development as there is even more overfeeding than in the tree. It is much faster to pick up parameters via reinforcement, although with overfeeds too, but in seconds. There are no constant patterns in the time series in the market, they should be taken from adjacent correlated BPs, and then it will be more stable. In fact, everything is very simple, but everyone is doing the wrong things. But to understand it and to feel the futility and hopelessness it is necessary to try. In any case, the more people the better, but not floodriders :)
 
Aleksey Vyazmikin:

Don't be sad about other people's stupidity - get busy and enjoy your achievements!

I do not understand people who only say things are bad but do nothing to change the situation, even when they are openly offered to do so.

Bless you too, I have never understood people who try to discuss personalities at every opportunity - Wiki helps "Ad hominem".

Do you have low self-esteem? ...

Well, as for the essence of your message to me - once again, what did you offer? I wrote my opinion on your first sentence, do you want to talk about subsequent ones too? - What do you have there:

Aleksey Vyazmikin:

Yes the point is that there is often a breakthrough in one direction, asymmetrical development in one direction, and pooling knowledge would improve things many times over.

Aleksey Vyazmikin:

It seems to me that NS would work fine with a stationary structure because of its mathematical nature (function fitting), but the market, if anything, is stationary in infinity. That's how I see it now...

I don't want to be rude to people I don't know - my upbringing doesn't allow it... here are your posts where I see your fantasies and?

oh yes, you are trying to change the situation, you are assembling a team of professionals, you know how it looks from the outside: you have recruited smart and talented and let's go.... You go over there, you do this, .... and I'm gonna go around in circles.

)))))

 
the man is doing something similar https://smart-lab.ru/blog/353092.php
ДАТАМАЙНИНГ(Rapid Miner & R) УМЕНЬШАЕМ ПАРАМЕТРЫ РОБОТА
ДАТАМАЙНИНГ(Rapid Miner & R) УМЕНЬШАЕМ ПАРАМЕТРЫ РОБОТА
  • smart-lab.ru
В мире полу-мистического граалестроение бытует несколько устоявшихся аксиом. Авторитетные гуру внушали их на протяжении многих лет, как заботливые родители, детям, дабы обезопасить «нерадивых» от лишних шишек. Одним из таких утверждений является то, что количество параметров должно быть минимальным, а лучше, чтобы их не было совсем. Статья...