Gathering a team to develop an IO (decision tree/forest) in relation to trend strategies - page 13

 
What I think is that all your standard models don't work, not because the models are standard, but because you can't prepare a representative sample for your target. Roughly speaking, your data is crap and no machine learning methods can handle this crap. And if the training sample is representative, i.e., it is set up as much as possible before training, then the standard methods will work as it is, without complicated training algorithms.
 
At first I tricked my way out to upload the set of predictors in a clever way, then I removed unnecessary predictors, then I increased the representativeness of the sample at the categorical level, which increased the quality of training by about ten percent, which I sometimes lacked. I trained the model with categories and looked at how it reflects the real data and if the real values are OK, I put the model on the real one and that's what Misha with Bart Simpson in a Superman suit on his avatar is telling me. You must have forgotten me by now....
 
Gramazeka1:
My guess is, all your standard models don't work, not because the models are standard, but because you can't prepare a representative sample for your target. Roughly speaking, your data is crap and no machine learning methods can handle this crap. If your sample is representative, i.e. best fitted before training, then standard methods will work just fine, without fancy learning algorithms.

What do you mean it doesn't work? Doesn't bring in 99% of the correct trades? I would like to hear the criteria for not working. The aim is to improve results, through specialisation.

Gramazeka1:
First I made a sly twist to load all predictors in a smart form, then I removed unnecessary predictors and increased the representativity of sample at the categorical level, which increased the quality of learning by about ten percent, which I was missing sometimes. I trained the model with categories and looked at how it reflected the real data and if the real values are OK, I put the model on the real one and that's what Misha with Bart Simpson in a Superman suit on his avatar is telling me. You must have forgotten me by now....

Have you lost access to your account?

 
Aleksey Vyazmikin:

What do you mean it doesn't work? Doesn't bring in 99% of the correct trades? I would like to hear the criteria for not working. The aim is to improve results, through specialisation.

Have you lost access to your account?

No, I'm just on my phone right now.
 
Gramazeka1:
No, I'm just using my phone right now

Why can't I log in to my account from my phone?

 
If you are constantly looking for new ways to learn, build and approach AI building, then I suppose the data for learning is not very good. With the pre-processing and amplification in R of the training data, Reshetov's optimizer works wonders. Unfortunately the elm networks from the doc showed their weaknesses, so I went back to Reshetov and strengthened the level of data preparation before training. The result is verified by the battle, but the quality of the training has increased as well. Increased the number of examples, within the limits of training quality and it seems to be... The same, but I have not yet understood and applied everything in the package. Two significant points so far. This is sifting out unnecessary predictors and increasing reprezenttivnost through categorization. Next we'll see what other tricks are in this package. I practically finished it yesterday evening on my knees. In general, let's see how and what happens next...
 
Aleksey Vyazmikin:

Why can't I log in to my account from my phone?

Because I don't remember my password and my email is authorised on my computer. Anyway...
 
Gramazeka1:
If you're constantly looking for new ways to learn, build and approach in building AI, then I suppose the data is not very good for learning. With the pre-processing and amplification in R of the training data, Reshetov's optimizer just works wonders. Unfortunately the elm networks from the doc showed their weaknesses, so I went back to Reshetov and strengthened the level of data preparation before training. The result is verified by the battle, but the quality of the training has increased as well. Increased the number of examples, within the limits of training quality and it seems to be... The same, but I have not yet understood and applied everything in the package. Two significant points so far. This is sifting out unnecessary predictors and increasing reprezenttivnost through categorization. Next we'll see what other tricks are in this package. I practically finished it yesterday evening on my knees. Anyway, let's see what happens next...

Of course, the sample has to be large, and you were right to increase the number of examples. What principle was used to sift out "unnecessary" predictors? Tell us about the categories, what do you mean by that?


Gramazeka1:
Because I do not remember the password, and e-mail is authorized on the computer. Anyway, not the point...

You don't know who someone is, there's a lot of that going around... that's why it's important to understand the who's who.

 
Aleksey Terentev:
I can help with deep python neural networks (picture), reinforcement learning, and linking all this stuff to mt.


As for the project management service, I plush for trello.com. And also as an alternative, ru.yougile.com.
You should familiarise yourself with the terms Kanban and Adjail.

Do you have any literature on RL for continuous tasks that is new?

 
Aleksey Terentev:
I can help with deep python neural networks (picture), training with reinforcement, and linking all this stuff to MT.

Neural networks - interesting. Not interested in trees/forests at all?

I just don't get it, do you wish to get on the team being created, or are you suggesting an alternative?


Aleksey Terentev:

As for the project management service, I'm in favor of trello.com. And also as an alternative ru.yougile.com.
You may want to familiarize yourself with the terms Kanban and Adjail.

Thanks, I had a look at yougile presentation - it seems very similar to social networks - convenient for current tasks, but I have not seen, where it is possible to store information, the result of intellectual work, so to speak.