Machine learning in trading: theory, models, practice and algo-trading - page 1697

New comment
 
Sadly, the moose has crept up unnoticed and without a twist. I'm going to a monastery :-(
 
Aleksey Vyazmikin:

It's beautiful. But I'm curious to see the grid breakdown of the predictor ranges that are subsequently enumerated.

Apparently just a grid on

Border count

The number of splits for numerical features.

The default value depends on the processing unit type and other parameters:

  • CPU: 254
  • GPU in PairLogitPairwise and YetiRankPairwise modes: 32
  • GPU in all other modes: 128
either evenly on a scale or by number of samples (percentiles)
 
Mihail Marchukajtes:
I wonder why you don't use vtreat for R? It just identifies levels in the input data relative to the target and thus selects those predictors which are significant for the target. In addition to classification there is an option for prediction. To be honest I don't know what I would do without it....

I do not know the principle of selection of predictors with this tool, so I do not use it.

Do you know this principle?

 
elibrarius:
Apparently just a grid on
either evenly on a scale or by the number of examples (percentiles)

If so, that's not good enough.

So far my comparison experiment is not finished - I found a mistake in data preparation and had to put 8 samples again for training.

I think that by the end of the week there will be the result, which I will publish here.

And, I'll do my "enumerator" of conditional leaves on these predictors (by the method that I described earlier) - with the control of the results (different metric) at each step.

The whole market cannot be described by a single model, so I will look for stable regularities, combining them in a TS, and I am not confused by the fact that part of the "space" will remain undefined.

 
Aleksey Vyazmikin:

I don't know the principle of selecting predictors with this tool, so I don't use it.

Do you know this principle?

Well if you believe what Doc says and if I understand him correctly for the classification it determines whether the levels on the input variable are formed depending on the result of the target. The way I see it. For all ones, the target significant input tends to fall in the vicinity of levels, which can also be several. In other words, if the target has 100 ones, then at the appearance of one, the significant input tends to fall into some region of the neighborhood of the past data. The target has 100 ones, and the important input has 30 neighborhoods, but whenever a one appears in the target, the important input is sure to fall into one of the neighborhoods. And the more it gets there when one appears, the more important it is. There is essentially a comparison with respect to the total set. In other words the result of work of this package will be selection of those variables which came to their neighborhoods more often than others. I'm sorry if I explained it in a confusing way. I'm drunk. I must think I had no time to wake up in the morning and ran into a moose and got carried away :-)
 
I will add from the other side. The vtrite calculates whether the target forms levels in the input data. And in those data where there are more of these levels and they are more specific, those data are stronger. In general, the script for R can lay out. He is the nickname of Doc, I just wrote it for myself. In order to get the result in a format that is convenient for me. FORMAT OPTIMIZER RESHETOVA UUUHAHAHAHAHAHA .... There already dopilit yourself for yourself. Only be sure to tell me whether to lay out or not....
 
Mihail Marchukajtes:
Well if you believe what Doc says and if I understand him correctly for the classification it determines whether the levels on the input variable are formed depending on the result of the target. The way I see it. For all ones, the target significant input tends to fall in the vicinity of levels, which can also be several. In other words, if the target has 100 ones, then at the appearance of one, the significant input tends to fall into some region of the neighborhood of the past data. The target has 100 ones, and the important input has 30 neighborhoods, but whenever a one appears in the target, the important input is sure to fall into one of the neighborhoods. And the more it gets there when one appears, the more important it is. There is essentially a comparison with respect to the total set. In other words the result of work of this package will be selection of those variables which came to their neighborhoods more often than others. I'm sorry if I explained it in a confusing way. I'm drunk. I must think I had no time to wake up in the morning and ran with the moose :-)

I do not understand the phrase "seeks to get into any region of the vicinity of past data.

There is no need to get drunk - you'll get drunk with the market.

The correction is overdue - I do not remember that on M15 the ZZ ray was pulled down by Donciann's channel (price channel) - since yesterday I started to buy. In addition, the RGBI index somehow went to the correction, despite the strengthening of the ruble yesterday - the Central Bank did not place all the bonds at the next auction.

 
Mihail Marchukajtes:
I'll add on the other side. The vtrite calculates whether the target forms levels in the input data. And in the data where there are more of these levels and they are more specific, those data are stronger. In general, the script for R can lay out. He is the nickname of Doc, I just wrote it for myself. In order to get the result in a format that is convenient for me. FORMAT OPTIMIZER RESHETOVA UUUHAHAHAHAHAHA .... There already dopilit yourself for yourself. Only be sure to tell me whether to download or not....

Do it.

Doc was looking for another way to select predictors - by building a decision tree on genetics - we jointly tested it before his departure.

 
Aleksey Vyazmikin:

I do not understand the expression "tends to get into any area of the vicinity of past data.

There is no need to get drunk - you'll get drunk with the market.

This correction is overdue - I do not remember so long pulling the ZZ ray downward on M15 on Donchianna channel (price channel) - since yesterday I started buying. The Rb.ua index was also decreasing, despite the ruble strengthening yesterday - the Central Bank did not place all bonds at the next auction.

This means that when a 1 appears, the input variable is rubbed near some value, say, 100. If a 1 appears in the output, the input variable is already near 100, thus the history begins to form the levels where the input variable comes to more often when a 1 appears in the target, and those who come to their levels more often, are the cooler ones. After all, we have a target with a look into the future. And if the input has come to 100, you can expect "1", which we will learn about only at the next signal. So it's like this....
 
Mihail Marchukajtes:
... Sorry if I explained it haphazardly. I am drunk. I must think, I`ve been drunk since morning and I`ve got into a mood :-)
Well... I was just beginning to get to the heart of the MOA from your posts, and here's such a confession ...))

Although, I wonder, perhaps I will understand the MO from the point of view of a drunken man and create the first drunken NS in the world?)))
1...169016911692169316941695169616971698169917001701170217031704...3478
New comment
Reason: