Machine learning in trading: theory, models, practice and algo-trading - page 2491

 
Mihail Marchukajtes #:
And in response I will tell you what this event is, it is the change of the sign of the smile. Now I sit and watch it in real life, or rather watched on Thursday when the big players have changed the sign of the minus on the plus and after that began a fierce fall. So I think we should look at the root and everything will work even without neural networks, and if we give them this tool it would be the bomb. It's like Singer's calculating machine, never gets tired and never makes mistakes.

that's your fantasy and nothing more.

 
mytarmailS #:

this is your fantasy and nothing more.

No, they are confirmed by practice, but not by machine :-(

 
Aleksey Vyazmikin #:

Not a simplification, but an alternative that will work faster, and that allows you to batch use many computers to find a solution!

Yes I mean the benefit globally to others, to popularize so to speak.

go for it! I'm all for it!

 
And now translate it into practice my main task and there are required a knowledge of macros Excel, I hope this does not confuse anyone?
 
What am I, like, not smart and do not know this or that? Just a glitch of Excel under Wine on Ubuntu led to the crash of the main system that I could not even boot, so I do not practice on a working computer. But I need an expert DDE and Excel!
 
mytarmailS #:

Well, our Misha is doing the same thing, his "initial rule" is the "TC sequent" from which he trains his AMO

You are essentially doing the same thing (absolutely) just call things by different names...

But your approach is smarter because you can choose any "starting rule" and he's tied to one...

The new method I'm thinking about is to find such rules, which I called significant events, in automatic mode, then training a model for each event and the main thing is to take into account the influence of the past event on the current one, and then on the future. The influence is expressed in terms of probability waves that may be of different length and predictive power. Thus, a sample on each bar can be divided into samples of significant events using basic predictors and then they can be combined for working on each bar, thus creating a complex model.


mytarmailS #:

I'll tell more, I do the same thing, especially if I want to fit 100 million predictors into my AMO, it's just technically impossible without compression/narrow rule/TC rule

For this I invented a "knowledge base" so that not to keep 100 million predictors in one table AMO will itself call the right ones from the right folder at the right time, but it's all at the stage of reflection...

Is it possible to train on 100k predictors? I'm doing selection of predictors through quantization now, checking the predictive ability of each quantile of predictor, and usually managing to find patterns that improve learning. I plan to generate predictors derived from others on the fly and check them at once, if it makes sense in the resulting sample then save the setup separately and use as an additional transformation before feeding model data, like a layer in NS.

 
Mihail Marchukajtes #:
And in response I will tell you what this event is, it is the change of the sign of the smile. Now I sit and watch it in real life, or rather watched on Thursday when the big players have changed the sign of the minus on the plus and after that began a fierce fall. So I think we should look at the root and everything will work even without neural networks, and if we give them this tool it would be the bomb. Like a Singer's calculating machine, never gets tired and never makes mistakes!!!!

What do the tests show? It's enough to collect statistics, and then think about online integration.

I trade wonderfully with my hands - 70% of profitable trades, but the market takes its toll when there is an emotional breakdown...

 
Aleksey Vyazmikin #:

What do the tests show? After all, it is enough to collect statistics on them, and then think about online integration.

I trade wonderfully with my hands - 70% of profitable trades, but the market takes its toll when there is an emotional breakdown...

I have the same :-(

 
Aleksey Vyazmikin #:

What do the tests show? After all, it is enough to collect statistics on them, and then think about online integration.

I also trade wonderfully with my hands - 70% of profitable trades, but the market takes its toll when there is an emotional failure...

Tests do not show anything as there is no data collection, but if I can do it with my hands, then I can do it with a machine 100%... In real trading there are certain nuances of preempting a bar before the key decision. And the bar decision with a lag of 1-3 bars, when the smile shows not what is going now, but what will be the next bar or in one bar, etc. does not always help, but when after clearing your smile changes its sign from +100 to -250 on this clearing!!!!!

And you understand with confidence that next ONLY sales, and look any signal Sequence is interpreted simply to the bottom and that's all, the most effective tactic !!!!!!!

 
Aleksey Vyazmikin #:

The new method I am thinking about is a search for such rules, which I called significant events.

Well, that's right... And now imagine that: in order for a trend to start, a series/sequence of such "significant events" (S/As) should happen, and many S/As happened one month ago from the current price, for example...

Can you imagine?

And now -- >

Aleksey Vyazmikin # :

Does it make sense to train on 100k predictors?

-- Think about how many different sequences of ZSs you have to go through to find a working one, there are trillions of variants...

And that's impossible to do in a sliding window (tabular data), like you do now (and all others), if you take the price in a sliding window then 2 predictors are enough to describe it all, but it's useless.