Machine learning in trading: theory, models, practice and algo-trading - page 1385

 
Yuriy Asaulenko:
You are wrong. That's the only way to do it.
 
Well, okay. If you want to work with a scale that depends on price, that's your right.
 
Yuriy Asaulenko:
Well, okay. If you want to work with scale that depends on price, you're right.

if the goal is to build a short-term model - your approach is fine, as long as the number of samples increases so much that all your returns can be thrown out as uninformative

lack of informativeness can be determined just when all features start to correlate with each other, in this case, increasing the length of the training sample will lead to nothing

i'm just puzzled by the question of lifetime extending

 
Maxim Dmitrievsky:

I assume that if you break the price into levels, then you can calculate the average depth of history by level, from how the price came to it and how it left it

But again, this would introduce additional errors in training?

i did a ZigZag, that showed the time between the tops of ZZ, alas, the price is so non-stationary that even the repetitions of time duration cannot be detected

If you want to find a solution with preprocessing of price, try the same Renko chart, at least the time component will be gone and there will be discrete levels (the height of Renko bricks)

 
Igor Makanu:

But this will again introduce additional errors in training?

I did a ZigZag, which showed the time between the formation of the tops of ZZ, alas, the price is so non-stationary that even the repetitions of the duration of time cannot be detected

you generally need some pre-processing solution, the same Renko chart may be an option, at least the time component will be gone and there will be discrete levels (the height of Renko bricks).

It will, yes. In general this approach has more questions than answers for me yet

but you have to crawl over to it somehow

 

IMHO this is a great moment to finally realize that MO is not working and it's time to go back to indicators for near-market earnings or look for a job in the service industry.


 
Maxim Dmitrievsky:

But you have to crawl over to it somehow

Renco is not a problem, I have an indicator for MT4, but there should also be indicators for MT5? - to feed the MO as a predictor of the value of the indicator

If i've got the Renco indicator in MT4 but i really want to use MO, if i look for ready solutions i'll use Python ((

 
Igor Makanu:

Renco is not a problem, I have an indicator for MT4, but there should also be indicators for MT5? - to feed the indicator values as a predictor for the MO

My aim is to change the indicator values for the time being in Python, I have decided on many current issues, but I really want to do MO, if I look for ready-made solutions, all on Python ((

Renco is not quite the right thing... you still need to divide it up into many levels

For python I made a socket connector, but in the tester mt5 sockets don't work, they said that maybe they'll make it work

 
Maxim Dmitrievsky:

if the goal is to build a short-term model - your approach is fine, as long as the number of samples increases so much that all your returns can be thrown out as uninformative

lack of informativeness can be determined just when all features start to correlate with each other, in this case, increasing the length of the training sample will lead to nothing

i just wondered how to prolong lifetime

Long-term models will require thinning anyway, and the model won't change.
 
Yuriy Asaulenko:
For long term models you will need thinning anyway, and the model will not change.

thinning is to throw out half of the information again and coarsen the model