Machine learning in trading: theory, models, practice and algo-trading - page 2354

[mytarmailS]

elibrarius:
It's better to compare profits. Not an error in inclination.

Not at all better, I don't see any advantage of watching profit, but I see a lot of disadvantages ...

I gave you the code, but there's no one to try it, it's much easier to write posts...

sibirqk:
In fact it's almost the same as building a trend line and then removing it from the original series. Yes such a residual is easier to predict, but it all comes down to predicting the trend. To forecast the trend it will be necessary to know approximately where the price will go in the future. But if we know that, then what for do we need all the previous stages.

How can you confuse detrending with normalization?

Ideologically it is closest to the Box-Cox conversion

[sibirqk]

mytarmailS:

Not at all better, I do not see any advantages to watch the profit, but I see a lot of disadvantages ...

I gave the code, but there is no one to try, it's obviously easier to write posts...

How can you confuse detrending with normalization, I can't even wrap my head around it...

Ideologically it's closest to the Box-Cox conversion

Well, you know better than that. I guess. Good luck with your research.

[Maxim Dmitrievsky]

Normalization\detrenders\smoothing/COC removes the last thing that was in the price (alpha)

[sibirqk]

Maxim Dmitrievsky:
Normalization/Detrenders/Smoothing/COS removes the last thing that was in the price (alpha)

Here I think I agree - to find alpha, imho, you need to learn to predict remote 🙂 .

[Maxim Dmitrievsky]

sibirqk:
Here I think I agree - to find an alpha, imho, you need to learn how to predict remote 🙂 🙂 This goes against the classical training of neural networks, which like to train homogeneous data.

This goes against classic data training for neural networks, which like to learn from homogeneous data

[sibirqk]

Maxim Dmitrievsky:

This goes against classical data preparation for neural networks, which like to learn from homogeneous data

Maybe that's why few people find alpha🙂

[mytarmailS]

blah - blah - blah - blah - blah

Why do something when you can just talk about it...

[Forester]

Has anyone figured out what fractional differentiation is?

At https://dou.ua/lenta/articles/ml-vs-financial-math/ he got it from the Prado.

He writes that"The time series differentiation that we know removes all memory of price evolution" - apparently, if for each bar we take the difference from the previous bar.

Here on this forum, most of them use the difference from the 0th bar.

1) And fractional differentiation - what is it? Coefficients of 0.1-0.5 are recommended.

You can't take a difference of less than 1 bar. Maybe it is a difference of 2, 5 ... 10 ... 20 bars from the next?

2) How is it better than the 0-bar difference?

[Maxim Dmitrievsky]

elibrarius:

Has anyone figured out what fractional differentiation is?

At https://dou.ua/lenta/articles/ml-vs-financial-math/ he got it from the Prado.

He writes that"The time series differentiation that we know removes all memory of price evolution" - apparently, if for each bar we take the difference from the previous bar.

Here on this forum, most of them use the difference from the 0th bar.

1) And fractional differentiation - what is it? Coefficients of 0.1-0.5 are recommended.

You can't take a difference of less than 1 bar. Maybe it is a difference of 2, 5 ... 10 ... 20 bars from the next?

2) How is it better than the 0-bar difference?

https://www.mql5.com/ru/articles/6351

I don't see much difference with detrend by EMA, and if you pass several rows with different lag, then the point of using fractional differentiation disappears.

[Maxim Dmitrievsky]

Then you will have questions about metamodels, and then the book. But I hasten to disappoint you - they also do not improve the results :D