Machine learning in trading: theory, models, practice and algo-trading - page 2364

[mytarmailS]

mytarmailS:

Find a solution...))

predictable...

[Maxim Dmitrievsky]

already wrote how to remove serial correlation at sliding window almost to zero, when preparing data

I did not play with scale invariance of signs/patterns, one would think. It depends on what to take as a reference point

The simplest one is to change attributes (for example, MAC periods) when volatility changes, etc. There will be a similar effect.

[mytarmailS]

Maxim Dmitrievsky:

already wrote how to remove serial correlation at sliding window almost to zero, when preparing data

I did not play with scale invariance of signs/patterns, one would think. Depends on what to take as a starting point.

Oooh this is one of the many questions to think about...

==========

About the solution to the problem itself...

I generate rules with a grammar regression (genetics), each rule can be considered an event, no binding to indexes, just a sequence of triggers

Something like this.

 [1] "SMA30 >= SMA10 & SMA60 <= 10.7 & SMA10 <= 16 & LO >= CL & SMA60 >= -11.5 & CL >= -3.8"                                                               
 [2] "LO <= 10.7 & SMA30 <= 16.4"                                                                                                                          
 [3] "OP >= -10.3 & LO <= SMA30"                                                                                                                           
 [4] "SMA30 <= HI & HI >= 18.8 & LO >= HI & SMA30 <= LO & LO <= CL & OP <= SMA60 & HI >= 6.7 & HI <= -2.6 & CL <= SMA30 & HI >= -19.2 & SMA10 >= SMA30"    
 [5] "SMA30 >= SMA60 & SMA60 >= -11.9 & SMA60 >= 10.3 & CL >= -4.2"                                                                                        
 [6] "LO <= 18 & CL >= -9.5"                                                                                                                               
 [7] "SMA10 >=  0.6 & HI <= -18 & LO >= SMA10 & LO <= -18.8"                                                                                            
 [8] "LO <= OP & LO <= 13.1"                                                                                                                               
 [9] "OP >= -20 & CL >= LO & LO <= -19.6 & HI >= -18.4"                                                                                                    
[10] "SMA30 <= 6.7 & CL >= -17.6 & CL <= -20 & HI >= HI & OP >= LO & LO >= LO & OP <= -19.2"                                                               
[11] "HI >= HI & SMA30 >= SMA10"                                                                                                                           
[12] "SMA60 <= 10.7 & SMA10 <= 16 & LO >= CL & SMA60 >= -11.5 & CL >= -3.8"                                                                                
[13] "SMA60 <= OP & SMA60 >= -9.5 & SMA60 <= 9.1 & SMA30 <= OP"    
....
..
..

If all rules are fulfilled at the segment, check if there was a bounce or what you want (whatever you put in the fitness function is what will happen)

These rules really hold regularities and have lifetime of 100-200 hits

Rules can be combined into ensembles (something like Random Forest)

But all this is terribly resource-intensive, but very promising, you can use absolutely any mathematics and architecture, and any target through fitness

You can try everything if you have the computing power.

[Maxim Dmitrievsky]

mytarmailS:

Oooh this is one of the many questions to think about...

==========

Regarding the solution to the problem itself...

I generate rules with a grammar regression (genetics), each rule can be considered an event, no indexes, only a sequence of triggers

It's something like this.

If all rules are fulfilled at the segment, check if there was a bounce or what you want (whatever you put in the fitness function is what will happen)

These rules really hold regularities and have lifetime of 100-200 hits

Rules can be combined into ensembles (something like Random Forest)

But all this is terribly resource-intensive, but very promising, you can use absolutely any mathematics and architecture, and any target through fitness

You can do anything, if you had the computing power.

the same as with seasonal or other filters, only the conditions are more complex

it is good for MI, because groups of examples are sometimes similar

[Aleksey Vyazmikin]

mytarmailS:

Fourier, I take the sum of the first n (2-5) harmonics with the largest amplitude

I.e. is it adjusted by eye to the plot?

[Aleksey Vyazmikin]

mytarmailS:

You can search everything at all, if you have deductions.

Is this method faster than the genetic tree I posted earlier or not? The outcome I see is the same - obtaining leaves and combining them into groups - I've already done all this.

I can calculate what is required to get intermediate results for their evaluation.

Added: True, there's also a rule here for comparing one indicator to another - that's really new - I've been thinking about this for a long time.

[Maxim Dmitrievsky]

Evgeny Dyuka:
the only useful thing that Microsoft has done for the world is VSCode

The python decorators can not fix it interactively, I am tired to send them bug reports

The only useful thing Microsoft did to the world was the VSCode python decorators can not fix it interactively.)

[Aleksey Nikolayev]

Maxim Dmitrievsky:

Python decorators can't fix it in interactive mode, I got tired of sending them bug reports

I've sent them bug reports from two computers and people have written to me.)

So, it is time to switch to R)

[mytarmailS]

Aleksey Vyazmikin:

I.e. is it adjusted by eye to the site?

differently.

Aleksey Vyazmikin:

Is this method faster than the genetic tree I posted earlier or not?

No it is not, it is fundamentally different rules...

Methods can not be compared"genetic programming" is a direction of methods in which some programs write other programs, I just implemented it in the form of rules (can be anything)

genetic tree is a special case of design tree with a flavor of gen. algorithm. i.e. the usual broken hat because the input goes to the X,Y with a binding to the indices

Aleksey Vyazmikin:
Added: True, there's also a rule here comparing one index to another - this is really new - I've been thinking about it for a long time.

there could be anything

[Aleksey Vyazmikin]

mytarmailS:

because the input is X,Y linked to the indices

Whatever you put in, that's what's going to happen.

Okay, I get it, you don't need any help.