Machine learning in trading: theory, models, practice and algo-trading - page 941

 
Maxim Dmitrievsky:

I'm not sure, there's not a lot of information on this.

But it turns out, for example, that if in the last quarter of the year I boo a model, it works fine for the whole year, and then it breaks

something like that...

if short-term, then it works for about 3 months, and then it breaks ... i.e. again we get into a cycle, but quarterly

Well, let's say it breaks, and then in 9-12 months it does not start working normally again?

 
Aleksey Vyazmikin:

Well, let's say it breaks, and then 9-12 months later it doesn't start working properly again?

No, it will never work again.

The patterns are only inside the cycles, outside they are different

 

I.e. it turns out it is necessary to train at least on the 1st semiphase of each new cycle, or as it is called scientifically

if it has not yet passed, then on a larger and more positively, or on a smaller

it is a theory supported only by small observations and common sense :)
 
And where, in fact, is Alyoshenka the son with his 100% accuracy? Help us, save us suffering - put the Grail right here. For it will be rewarded.
 
Dr. Trader:

Maybe combining all this information into one tree will outweigh the disadvantages of using three classes, and the puzzle will come together :)

I put everything into one table by this principle.

//--Минимум
      if (finRez_Buy>=0)Klass_Buy=2;
      if (finRez_Sell>=0)Klass_Sell=-2;
//--Максимум
      if (finRez_Buy>PoiskProfitMax)Klass_Buy=1;
      if (finRez_Sell>PoiskProfitMax)Klass_Sell=-1;
//--Фильтер
      if (finRez_Buy<0)Klass_Buy=3;
      if (finRez_Sell<0)Klass_Sell=-3;
      
if (Klass_Buy==1 && Klass_Sell==-1)Klass=-3;
if (Klass_Buy==2 && Klass_Sell==-2)Klass=3;
if (Klass_Buy==3 && Klass_Sell==-3)Klass=3;

if (Klass_Buy==1 && Klass_Sell==-2)Klass=1;
if (Klass_Buy==2 && Klass_Sell==-1)Klass=-1;

if (Klass_Buy==1 && Klass_Sell==-3)Klass=1;
if (Klass_Buy==3 && Klass_Sell==-1)Klass=-1;

if (Klass_Buy==2 && Klass_Sell==-3)Klass=2;
if (Klass_Buy==3 && Klass_Sell==-2)Klass=-2;

      arr_Buy[i]=Klass;

I also grouped them by arr_TimeH predictor - maybe this way will help me in some way.

I attach the files.

Files:
New.zip  3807 kb
 
Maxim Dmitrievsky:

I.e. it turns out it is necessary to train at least on the 1st semiphase of each new cycle, or as it is called scientifically

if it has not yet passed, then on a larger and more positively, or on a smaller

this is a theory backed only by a little observation and common sense :)

Yeah, only usually when it becomes obvious that the new phase has arrived already it, the new one, comes to an end....

Or there is a reliable way to identify it?

 
Alexander_K2:
And where, in fact, is Alyoshenka-son with his 100% accuracy? Help us, save us who are suffering - post the Grail right here. For it will be rewarded.

Uncle Sashka, come out of the Middle Ages!

People have been using the matlab for a long time now)

At the bottom there's information about metatrader, matlab, increments, and neuronics.

 
Aleksey Vyazmikin:

Yeah, only usually when it becomes obvious that the new phase has arrived it, the new one, comes to an end....

Or is there a reliable way to identify it?

Well, quarterly, within a quarter there is less probability of big changes in the conjuncture. Make a decomposition of quotes over a long time period and see. Not Fourier but something more interesting, like a mode empirical. R is easy, I guess. Probably can improve stability and survivability with this.

need more mana, in short )

The link to decomposition is very well inserted, I've already used it, it's cool. But I haven't tried it within the framework of such studies of cycles. I will do it tomorrow.)
 
Renat Akhtyamov:

Uncle Sashka, come out of the Middle Ages!

People have been using the matlab for a long time now.)

At the bottom there's information about Metatrader and Matlab, about increments and neuronics.

Thank you. He will read it later. Right now he's asleep, reading some Shelepin. He told me not to disturb him.
 
Maxim Dmitrievsky:

Well, quarterly, within a quarter there is less probability of large changes in the conjuncture. Make a decomposition of quotes over a long period of time and see. Not Fourier but something more interesting, like a mode empirical. R is easy, I guess. Probably can improve stability and survivability with this.

Need more mana, in short )

The link to decomposition is very well inserted, I've already used it, it's cool. But I haven't tried it within the framework of such cycle studies. Tomorrow I will do it.)

I still haven't made friends with R, so I'd be interested to see what you come up with!