Machine learning in trading: theory, models, practice and algo-trading - page 2917

New comment
 

Some kind of life hacks that allow you to algorithmise the unalgorithmable,

turn into code what can't be explained, what the eye sees, what the brain understands, but you can't describe it in code...


For example, we have a price.

Traders with enough experience can mark the levels of the BSP on the chart, often with quite good accuracy,

like this.

That is as if they separate the beginning of another movement from the first movement, our eye sees it, our brain understands it, but writing an algorithm is out of the question, it will always be incomplete, not correct, not working... It is the same story that a successful manual trader cannot write his TS as an algorithm, there is always inaccuracy and incompleteness in explanations (even to himself).


But if we prepare the data correctly and cluster the data correctly, we will get something similar to what the human brain sees on the chart....


As you can see, the algorithm has divided prices into zones/clusters/states...

coloured zones are clusters, white zones are what the algorithm considers noise....

And now the most interesting thing, what are the PD/SP levels in terms of the second picture above? They are transition zones from one cluster to another (from one state to another).


We can't see it, but we intuitively understand it, and now it's possible to algorithmise it.


And I forgot to say, it always works, not just in this one picture.

 
mytarmailS #:

Some sort of life hacks that allow you to algorithmise the non-algorithmisable,

to turn into code what cannot be explained, what the eye sees, the brain understands, but you can't describe it with ordinary code....

...

But if you prepare the data correctly and cluster the data correctly, we will get something similar to what the human brain sees in a graph....

"Properly" how? What kind of clustering? How do you transition from state to state on one bar, because to identify a cluster you need some information about the formation of the cluster?

 
Changed my name from elibrarius to woodsman, since I'm doing wood modelling... to make the nickname match the main activity.
 
Aleksey Vyazmikin #:

"Right," how? What kind of clustering? How do you transition from state to state on a single bar, because cluster identification requires some information about the formation of the cluster?

wrote in PM

 
Forester #:
Changed my name from elibrarius to forester, since I'm doing wood modelling... to make my nickname match my main activity.

There's a nice package with the same name forester. You might be interested.

 
I need to rename myself R_fanat then))))
 
Vladimir Perervenko #:

a nice package with the same name forester. You'll be interested.

Well, I have my own forest. The advantage is that you can change absolutely anything in the code for experiments. So I don't look at other people's forests/busts. Unless something breakthrough appears. But trees are simple, it's hard to come up with something new.
 
Forester #:
Well, I have my own forest. On the plus side, you can change absolutely anything in the code to experiment.

What's the use?

All the necessary parameters are already in the packages and you can change them.

 
mytarmailS #:

What's the use?

All the necessary parameters are already in the packages and can be changed.

It is of little use (for trading), but there is more room for experimentation.
 
From John Simons Renaissance
The team improved their predictive algorithms by applying a fairly simple quantitative metric - how many times a company was mentioned in a news feed - regardless of whether the mentions were positive, negative or just rumours.
1...291029112912291329142915291629172918291929202921292229232924...3595
New comment