Machine learning in trading: theory, models, practice and algo-trading - page 3489

 
please moderators, disregard, this is a long time ago with this long time member. interesting.
 

I have trained my method on random data with peeking on two samples, just like here with normal sampling.

Well, let me say that it is quite trainable. Learning dynamics - probability of choosing class "1" at each iteration after "training".


This is what the probability of choosing the correct quantum segment at each iteration looks like. It's shaky at the end as the sample decreases.

Overall very similar dynamics, as in the data we consider not random.

And this is what the graph looks like already with the number of good options to choose from.

The difference in the peak is about 4.5 times with non-random data. Can we then consider that 20% on non-random data are just randomly selected quantum segments? I note that CV without mixing is used here for selection - the sample ranges are split into 10 parts and they are subject to estimation in the aggregate.

And who is interested in the balance on the sample - here it is - 1 point 0.00001, the spread is taken into account.

As a result, we took a random sample, we got lucky with some probability, built a random model or tweaked the variables in the optimiser, and now we are happy with the result and carry our hard-earned money to the market....

 
Aleksey Vyazmikin #:
Alexei, how many cores or threads do you have on your server?
 
mytarmailS #:
Alexei, how many cores or threads do you have on your server?

I have many computers, max 28 cores - 56 threads - 128 RAM.

 
Aleksey Vyazmikin #:

I have a lot of comps, max 28 cores - 56 threads - 128 RAM.

Do you use them? 56 threads.
 
mytarmailS #:
Do you use them? 56 threads

Recently I do not use all resources so often - I was engaged in theoretical developments on one sample, but soon I want to test the developments on different samples that would evaluate the result more objectively, for this purpose it will be convenient to use a separate computer for each sample - if to train models on the same CatBoost.

Also for the EA from the market I used all resources - for selection of basic tables of parameter settings.

Initially under selection of leaves all bought and under genetics - here can still write a code under idea on python, that used on R - has disadvantages, which I did not understand how to eliminate.

In general, where there are a lot of calculations and my algorithm seems useful - I transfer the logic to OpenCL and calculate on a video card.

 
Aleksey Vyazmikin #:

I have a lot of comps, max 28 cores - 56 threads - 128 RAM.

If I throw you a script, can you calculate on all cores, if anything?
 

Gemma


 
mytarmailS #:
If I throw you the script, can you do the calculation on all cores, if anything?

I think you need to figure out at the beginning how long it will take.

What kind of processor do you have?

 
Maxim Dmitrievsky #:

Gemma


Well, well - sounds like humour, curious....