Machine learning in trading: theory, models, practice and algo-trading - page 2629

[СанСаныч Фоменко]

Aleksey Nikolayev #:

It is possible to understand the word "seeking" in this context in different ways. At least in two ways. One can either try to manually find signs that catch the "physics" of the market, or one can try to construct them by means of MO from raw prices.

Searching means comparing values of predictive power.

But "catching" is closer to a dating site.

[mytarmailS]

Maxim Dmitrievsky #:

a recommendation came up on Medium on your topic, it may come in handy, I didn't get into it

I was interested in this approach because the trained models can be easily transferred to the terminal (I think)

https://medium.com/@james_laidler/generating-a-rules-based-system-using-iguanas-762843dd1418

Not Max, that's the wrong coat

[Aleksey Nikolayev]

СанСаныч Фоменко #:

Searching means comparing values of predictive power.

You can't compare something that isn't found.

SanSanych Fomenko #:

But "hooking" is closer to a dating site.

I won't argue - everyone has their own associations.

[Maxim Dmitrievsky]

From my observation, the sliding feature window is not an absolute evil, as it works for predictable series

The absolute evil is over-differentiating on the weakly predictable, which kills information about price levels, which themselves carry some information

But machine learning is used to working only with stationary features, with the exception of linear regression

[mytarmailS]

Maxim Dmitrievsky linear regression

Both are evil.

And levels have to be remembered and invariant to time

===========

By the way, you had an interesting idea about copying the TS from a Market. It is naive, but it can be a criterion of intellectuality of the algorithm for searching for solutions. If the algorithm can guess the rules of TS, it means it knows something... In general, it is useful to discuss the question: What algorithm should be able to decode the TS of others?

[Maxim Dmitrievsky]

mytarmailS #:
Both are evil.

And levels must be remembered and invariant to time

===========

By the way, you had an interesting idea about copying TS from the Market. It's a naive idea, but it may be a criterion for intellectuality of the search algorithm. If the algorithm can figure out the rules of TS operation, it means it knows something... In general, it's useful to discuss the question: What should be this algorithm to be able to decrypt other's TS

You just need to download the MT5 tester reports, they are in xlsx format. There, everything, time and profit trades of any bot from the Market (you can test any paid bot for free, then save the report).

but there is a lot of unnecessary information like tables and graphs on top.

i.e. first, make a convenient bot report download, and then think about the learning algorithm

The irony is that there are not many worthwhile bots in the Market, they also pour in after a while

[mytarmailS]

Maxim Dmitrievsky #:

You just need to download the MT5 tester reports, they are in xlsx format. There everything, time and profit of any bot from the market (you can test any paid bot for free, then save the report)

but there is a lot of unnecessary information like tables and graphs on top.

i.e. first, make a convenient bot report download, and then think about the learning algorithm

the irony is that there are not many bots in the Market, they also pour in some time

Yes, to spar is nonsense, my message is to think of an algorithm that can solve the TS, you understand that just to train a MOSHKA will not work

1. We need to figure out how the algorithm will remember levels

2. How will build rules based on both indexes and events (without indexes)

Automatic creation and enumeration of billions of attributes

4. Solving TC is not an approximation, not an approximation to the result, it is a clear solution, like the password in Enigma. This is a different paradigm

[Forester]

mytarmailS #:
I'm just trying to think of an algorithm that could solve the TS, you understand that just teaching the MO won't work.

If there were a million examples then maybe MO could average examples to any logic or combination of indicators etc. If there were 1000 examples, it would already be something very averaged and unlikely to match the original EA.

[Maxim Dmitrievsky]

mytarmailS #:
Yes sparsity is nonsense, my message is to think about an algorithm that could unravel the TS, you understand that just teach the MOSHka will not work

1. You need to figure out how the algorithm will remember the levels

2. How to build rules based on both indexes and events (without indexes)

Automatic creation and enumeration of billions of attributes

4. Unlocking the TS is not an approximation, not an approximation of the result, it is a clear solution, like the password in the enigma. This is a different paradigm

It's kind of creative, you never know.

The trades of a profitable TS should point to a pattern. If it only uses price/time, then picking/approximation seems possible

[Forester]

I think the best way is to run the indicator in the tester on all possible instruments for all time, and on a custom symbol generated by the GSC. Collect millions of trades on all possible combinations of the chart and train on it. This is the only way to get close to any logic inherent in the Expert Advisor.