Machine learning in trading: theory, models, practice and algo-trading - page 944

 
Aleksey Vyazmikin:

Yesterday I also made a set for 2017 - there is also within 30% of the correct data, but what surprises me on different (divided into two groups of predictors) sets of predictors I get similar results. I'm thinking, is it randomness, or just an opportunity to use the forest?

There is no need to look at accuracy, it is a bad metric, especially when there are a lot of classes. For example, in the case where I trained the last tree, 50% of all examples are "0" class. The tree could always return "0" at all, and already be right 50% of the time (instead of 33% according to random), the result seems to be better than random, but it is obvious that the tree is useless.
A better metric is for examplehttps://en.wikipedia.org/wiki/Cohen's_kappa


Aleksey Vyazmikin:

Merging is better to do from the last attached file - there division is slightly more accurate than 0 - all that did not make a profit, and in earlier files 0 - this and that brought a very small profit.

It's better to do it yourself. There are two targetets, each with a bunch of classes, I don't want to get confused.
I need one targetet with 3 classes -
"-1" to enter short.
"1" for entering long.
"0" if long and short positions are both negative (or both are positive, if there is such a case).
This tree is easier to teach than the previous 4. It's easier to use it to trade.


Aleksey Vyazmikin:

However, what surprised me was that by increasing the target classification worked better outside of the training sample. Maybe the opposite should be true, increase the number of different classifications?

It should be checked many times. If it is confirmed it is cool. Maybe, the tree will learn to identify certain classes well, for example, the profit of at least 100 points or something similar. Then it will be possible to make new targets with such logic for the real account.
For example, there should be 10 classes - "0" if there is a loss in any case, "1" if long position has a profit from 0 to 100 points. "2" if long profit is from 100 to 200 pips. "-1" if short profit is from 0 to 100 pips. "-2" if short profit is between 100 and 200 pips. So on. Then according to the summary table of correct and wrong answers you will be able to see if some class will be clearly better than others.
p.s. I've never done it myself. Maybe I'm not making sense. But if you're into trees and lots of classes, then this test is a logical continuation of the experiment.

 
Aleksey Vyazmikin:

PS I can give you access to a laptop on TW, that you would not burden your PC.

Now do not, I have a mess in several R scripts, trying different methods of selecting parameters. If something eventually fits, and the result is better than in your program - script attached here, you can continue to experiment.

 
Maxim Dmitrievsky:

You don't get the point, it's not at that level of abstraction, forget it.)

What does the level have to do with it, the question is exactly the subject. A regularity is a similar reaction to an external factor (irritant), in our case, for example, if the price went X points up, the probability that it will soon correct is higher than not. Or, every 5 bullish candlesticks in a row of small candlesticks indicates a high probability of a big bearish candlestick, which will outbid at least 3 of the bullish candlesticks. But if the probability of these events changes, then for me it is a change of patterns, and then the TS stops working.

But your vision is different?

 
Aleksey Vyazmikin:

Do you have a different vision?

I've already described my vision lol

And it's not just a vision, it's reality. It's up to everyone to accept it or not.

The level of abstraction is not a stick or a pickle, but why do the patterns change in the market and with what frequency, cause and effect. I am tired of repeating myself.

 
Dr. Trader:

The accuracy is not necessary to look at, it's a bad metric, especially when there are many classes. For example, in the case where I trained the last tree - 50% of all examples are class "0". The tree can always return "0" at all, and it will be right in 50% of cases (instead of 33% according to random), the result seems to be better than random, but it is obvious that the tree is useless.
A better metric is for examplehttps://en.wikipedia.org/wiki/Cohen's_kappa

Later I will pull all the rules from the two models and see what the result is, if you add the two models. Now put the network for the sake of interest, maybe something will pick up, but already picking for 12 hours - is it generally normal for networks?

Dr. Trader:

Better to do it yourself. There are two targetets, each with a bunch of classes, I do not want to get confused.

I need one targeting with 3 classes -
"-1" to enter shorts.
"1" for entering long.
"0" if both long and short positions are negative (or both are positive, if there is such a case).
This tree is easier to teach than the previous 4. And it is easier to trade using it.

Okay, I'll do it soon and post it here.


Dr. Trader:

This has to be repeatedly tested, if confirmed, then cool. Maybe the tree will learn to identify specific classes well, like profit at least 100 points, or something like that. Then it will be possible to make new targets with exactly this logic for the real account.

For example, we may open ten classes - "0" if long position will be lost in any case, "1" if long position will have profit from 0 to 100 points. "2" if profit to long position is from 100 to 200 points. "-1" if short profit is from 0 to 100 pips. "-2" if short profit is between 100 and 200 pips. So on. Then according to the summary table of correct and wrong answers you will be able to see if some class will be clearly better than others.
p.s. I've never done it myself. Maybe I'm not making sense. But if you're into trees and lots of classes, then this test is a logical continuation of the experiment.

I think that the tree works on elimination, i.e. if it finds 1, then all the rest 0, and when there are more options, it starts to classify the remaining zeros, rather than dumping them into one pile, which reduces the error (but this is my hypothesis). And to help it classify these other 2,3,4 I think to add a predictor that will indicate not only financial result from input, but on the basis of available other predictors will form a signal to input, then there should be a feedback, i.e. I think to add different justification of inputs and their financial result. The complexity here will be in the case of conflicting inputs for different TS, perhaps just put this situation in a separate group.


Dr. Trader:

I have a mess in several R scripts, I'm trying different methods of selecting parameters. If something will eventually work, and the result will be better than in your program - I will attach the script here, you can continue to experiment.

If I need it, please let me know, while the hardware is idle anyway.

 
Maxim Dmitrievsky:

I've already described my vision lol

And it's not just a vision, it's reality. It's up to everyone to accept it or not.

The level of abstraction is not a stick or a pickle, but why the patterns in the market change and with what frequency, cause and effect. I'm tired of repeating myself.

That's how I answered that the tool shows nonsense and focuses on trends, which is not true.

 
Aleksey Vyazmikin:

So I responded that the tool shows nonsense, accentuates the trends, which is not true.

agreed

 
Maxim Dmitrievsky:

agreed

I.e. the cycles may be there, just the tool is not the same - it is a visual perception.

You please tell me if he has 1001 criteria and I am mistaken. It makes no sense to argue, I said about my visual perception, if I'm wrong, I would like to know that I am wrong.

 
Aleksey Vyazmikin:

I.e. the cycles may be there, just the tool is not the same - it is a visual perception.

You please tell me if he has 1001 criteria and I am wrong. There is no point for me to argue, I said my visual perception, if I am wrong, I would like to know I am wrong.

I showed some actual periodicals, and decomposition defines them too

Vladimir Perervenko uses spectral analysis, which is very similar

How to use it in practice, let everyone to apply his sick creative imagination, because I'm not giving advice until the grail is finished

 
Maxim Dmitrievsky:

I have shown some real periodicities, and decomposition determines them too

Vladimir Perervenko uses spectral analysis, which is very similar to

How to use it in practice, let everyone use his sick creative fantasy, because I'm not giving advice until the grail is done.

Do you think that these cycles exist on small TF? Let's see, there will be more bars there!