Machine learning in trading: theory, models, practice and algo-trading - page 1827

 
On the topic of randomness of quotes and garbage on the input...

I made a comparison of methods for evaluating the importance of predictors. https://www.mql5.com/ru/blogs/post/737458

Even they are impossible to evaluate objectively and to sift out the important ones from the garbage ((
Сравнение разных методов оценки важности предикторов.
Сравнение разных методов оценки важности предикторов.
  • www.mql5.com
Провел сравнение разных методов оценки важности предикторов. Тесты проводил на данных титаника (36 фичей и 891 строки) при помощи случайного леса из 100 деревьев. Распечатка с результатами ниже. За образец с которым можно сравнивать все методы взял результаты оценки предикторов с реальным удалением предиктора/столбца из обучающего набора и...
 
elibrarius:
To the topic of randomness of quotes and garbage on the input...

I made a comparison of methods to estimate the importance of predictors. https://www.mql5.com/ru/blogs/post/737458

Even they are impossible to estimate objectively and to sift out important from trash ((

And the best way is to run through the models with new data.)

 
mytarmailS:

It's not ticks, it's 5minclose prices, it takes me a long time to draw candles in R

It is clear that kloz, but if the volatility in the direction of the lot is greater than the lot, then the expectation of the lot is greater than half.

 
Maxim Dmitrievsky:

Or better at once retraining of models and new data )

I took this as a sample for comparison. But this is even for 891 rows - 35 retraining is long (about a minute). And if there will be a million rows and 100 columns? You can go out and smoke for a week or two).

 
Valeriy Yastremskiy:

It's understandable that kloz, but if the volatility is more than a loser, then the expectation of a loser is more than half.

What the hell does that have to do with anything? Damn it...!!!!

I'm saying that the market should be analyzed differently, that we should look for patterns differently, I'm saying that the standard approaches are not applicable to the market and it's obvious, none of them works!

I gave an example of a pattern that none of the methods for working with time series will ever find, there are thousands of such patterns, this one is just for illustration ... !!!!!!!

And you're saying to delete the trend, or you're counting volatility, do you have any idea what we're talking about?

 
elibrarius:

This I took as a sample for comparison. But this is even for 891 rows - 35 retraining is long (about a minute). And if there will be a million rows and 100 columns? Then you can go out to smoke for a week or another )))

So there are many combinations, it will be faster in 10 years) That's why predictors are important, it would be possible to simply look through them all and choose the best one, but so far it's a long time.

 
mytarmailS:

What the hell does that have to do with anything? !!!!

I say that we need a different way to analyze the market, to look for patterns, I say that the standard approaches are not applicable to the market and it's obvious, none of them works!

I gave an example of a pattern that none of the methods for working with time series will ever find, there are thousands of such patterns, this one is just for illustration ... !!!!!!!

And you're saying to delete the trend, or you're counting volatility, do you even know what we're talking about and what's going on?

It's like this: we start with ticks, bars, and a very long series, which we scale. What are you suggesting? Really, it's not clear yet.

 
Valeriy Yastremskiy:

How else is it?, initially ticks, bars, a very long row, which we scale. What do you suggest? Really, it's not clear yet.

Synthesize rules that take into account strictly some specific events in a specific sequence, and also take into account the price at which these events occurred.

Here is a certain pattern.


This is a pattern in prices.

Here is the same pattern

 
mytarmailS:

synthesize rules that take into account strictly some specific events in a specific sequence, as well as take into account the price at which these events occurred.

There is nothing besides tick bars and volumes, the rest are derived from tick bars. To obtain the rules, you have to analyze BP. Is there anything else besides tick bars for rules?

 
Valeriy Yastremskiy:

Synthesize rules from what? There is nothing but tick bars and volumes, the rest are derived from tick bars. In order to get the rules, we would have to analyze BP. Is there anything else besides tick bars for rules?

What else do you need?