Machine learning in trading: theory, models, practice and algo-trading - page 1432

 
Ivan Butko:

So, it makes no sense to use 10 layers with 100 neurons each instead of 3 layers with 20 neurons each (as in the example)? I already wanted to load my cor i7 today.

I did a training with 10 layers of 100 neurons each on an hourly chart for a period of 1 year from 2018 to 2019. Out of 6,000 hourly candles it now guesses all 6,000. The most interesting thing is to assemble these 1000 strings into an EA and see if this "great" fitting will work for at least a month.

 
Ivan Butko:

Did the training with 10 layers of 100 neurons each on the hourly chart for a period of 1 year from 2018 to 2019. Out of 6000 hourly candlesticks now guesses all 6000. The most interesting thing remains - to gather these piles of strings into the Expert Advisor and see if this "great" fitting will work for at least a month.

What piles of strings? No, it won't work. The reason is non-stationarity of the market.

Try it on some simple function with addition of noise, then complicate it. If you have a lot of neurons then it will learn fine on any complex function, but on new data with more complexity it won't work.

Here's the whole thread about it, you can read from the beginning

 

Hey guys!

How's our AI bot coming?

 
Maxim Dmitrievsky:

What piles of lines? No, it won't work. The reason is the non-stationarity of the market.

Check on some simple function with addition of noise, then complicate it. If there are many neurons it will work fine on any complex function, but with new data, when complicating it, it won't work anymore.

Here's the whole thread about it, you can read from the beginning

there is a long line with weights - one neuron. To translate it from prog into proper form into EA, you need to change syntax and add a semicolon after each line. And there are 10*100 = 1000. It's boring, but what the article says, that's what I've been doing.)

Yes, I am aware that it will not work.) I read about retraining.

>>there's a whole thread about it, you can read it from the beginning.

1432 pages... D))
 
Renat Akhtyamov:

Does one trend prove anything?

But SB is not.

It all looks too staged - 300% by Fibo and a precise reversal by the "stick method" - someone's cunning plan...

 
Aleksey Vyazmikin:

It all looks too staged - 300% by Fibo and exact reversal by the "stick method" - someone's cunning plan...

300% is only 4 times more. There is no Fibo, Fibo is a ratio of 1.6 between numbers, i.e. 68% and other, these values can be called Fibo.

 
Grail:

Mmmm... that's pretty cool...

What did you use, if it's no secret?

catbust

 
Ivan Butko:

300% is only 4 times as much. There is no Fibo here. Fibo is a ratio of 1.6 between numbers, i.e. 68% and so on, these values can be called Fibo.

Yes, it's really not Fibo, it's even more primitive and brutal.

 
Aleksey Vyazmikin:

Who else believes in random walks?

Well you understand the arbitrariness of all these "trend lines", there are several obvious ones, some will "work", others will not, this is the main method of "proving" for gypsies and charlatans, shifting the sample in the favorable direction, choose examples where it works, ignore the rest. If you want to really prove something, then formalize the algorithm for drawing trend lines and market behavior in accordance with them, and then run it over a long history and collect statistics, that's the only way. And one picture, even 10 pictures, can easily be used to "prove" the superprofitability of any indicator or TS, it's a gypsy subject, don't get like that.

 
govich:

Well you understand the arbitrariness of all these "trend lines", there are several obvious ones, some will "work", others will not, this is the main method of "proving" for gypsies and charlatans, shifting the sample in the favorable direction, choose examples where it works, ignore the rest. If you want to really prove something, then formalize the algorithm for drawing trend lines and market behavior in accordance with them, and then run it through a long history and collect statistics, that's the only way. And one picture, even 10 pictures, it's easy to put together a "proof" of superproficiency to any indicator or TS, it's a gypsy topic, don't get like that.

You don't understand the main point, it's just the implementation of a plan for price behavior. Who says the plan has to be too complicated, and who says it has to be the same all the time?