Machine learning in trading: theory, models, practice and algo-trading - page 2121

 
Elibrarius:

I did, but it didn't work.

That's strange... Okay, I'll take another look at this thread - I'll let you know later.

 
By the way in the CatBoost code there is a method of quantization
GreedyMinEntropy

which is not mentioned in the official documentation.

 
Maxim, can you compare where the learning speed is faster, in python or in the console version of CatBoost?
 
Aleksey Vyazmikin:
Maxim, can you compare where the learning speed is faster, in python or in the console version of CatBoost?
The same, it's the same program
 
Alexander_K:

Strange... All right, I'll look at this topic again - I'll write back later.

At first I spoke out of recollection and was wrong.

Checked just now with cross validation. Time still gives 1-2% gain on new data. While the total preponderance of successful predictions over unsuccessful about 5% at TP = SL. Ie from 10 to 30% of the contribution to success comes from time.

However, this is a test on a 2-month piece of history. Things may change at another site.
 

I feed the time as cosine and sine. This was discussed here https://www.mql5.com/ru/forum/212376#comment_5983502

Simply as 0...5(day of week) or 0...23 (hour) or 0...59 (minute) can also be fed, but as a categorical variable.
Sine and cosine are already numeric, any algorithm will accept.

Обсуждение статьи "Глубокие нейросети (Часть II). Разработка и выбор предикторов"
Обсуждение статьи "Глубокие нейросети (Часть II). Разработка и выбор предикторов"
  • 2017.07.31
  • www.mql5.com
Опубликована статья Глубокие нейросети (Часть II). Разработка и выбор предикторов: Автор: Vladimir Perervenko...
 

there is a relatively new trend called active learning. It can self-sample the data in the best way. It seems to be suitable for my approach (random sampling). I haven't got into it yet.

https://libact.readthedocs.io/en/latest/overview.html

https://medium.com/towards-artificial-intelligence/how-to-use-active-learning-to-iteratively-improve-your-machine-learning-models-1c6164bdab99

Overview — libact 0.1.3 documentation
  • libact.readthedocs.io
libact is a Python package designed to make active learning easier for real-world users. The package not only implements several popular active learning strategies, but also features the active-learning-by-learning meta-algorithm that assists the users to automatically select the best strategy on the fly. Furthermore, the package provides a...
 
mytarmailS:

As I go along, the frequencies and possibly the phases are floating... The amplitudes are holding...

Here is the forecast for 500 points of the fitted model on the history of 10k of 4 harmonics

We can see that the forecast is relevant for all 500 points, but the frequencies are drifting, and they are drifting according to an incomprehensible algorithm

And this is an illustrative example, it may be even worse.

Have you tried it on the increments?

" One of the useful properties of a series of increments is the greater stability of the spectrum compared to the original process. ".This is Goodman writing.

 
elibrarius:

I feed the time as cosine and sine. This was discussed here https://www.mql5.com/ru/forum/212376#comment_5983502

Simply as 0...5(day of week) or 0...23 (hour) or 0...59 (minute) can also be fed, but as a categorical variable.
Sine and cosine are numeric, any algorithm will accept.

Can you post the function for converting time to sine/cosine? I would also try this method. In the article I published, the hour number was a significant predictor there. I wonder if this method is suitable for wooden models or more for neural networks.