Machine learning in trading: theory, models, practice and algo-trading - page 1616

New comment
 
Mihail Marchukajtes:
That's what I mean. Before referring to JPrediction I leave only 150 pieces out of 6000 thousands of columns, which are statistically significant, and only then I search them for that law describing the output. The number of columns should be twice as big as the number of table rows, theoretically speaking, so the algorithm has a lot to choose from. As a result, the optimizer leaves 5 to 10 pieces out of 150 suggested by me, which form the final model.

I noticed that of the historical features 3-7 work, the rest is garbage. For example, the number of the month, day of the week, etc. (2-3 pieces) will already work if there is repeatability. Only they have to be converted to categorical.

Increments and the like don't work. Or rather, they can work, but how - this is in the last article. It doesn't even have to be an MO for that.

yes, 5 to 10 is close, so it is to describe some kind of pattern
 
Maxim Dmitrievsky:

I noticed that of the historical features 3-7 work, the rest is garbage. For example, the number of the month, day of the week, etc. (2-3 pieces) will already work if there is repeatability. Increments and so on do not work.

I'll tell you this. The last version of Reshetov was version 14. I worked up the courage and completed the optimizer to version 15 and that was my personal decision. But what I added there changes the picture for the better. Proceeding from the theory that the most robust model is the one with minimal inputs and shorter polynomial, all other things being equal, I took a different path. If Reshetov's solution was model ascending, i.e. model ascending from 4 inputs and then adding more inputs so we end up with model with say 7 inputs, then I started with descending, so I start with 11 inputs and then decrease them so we end up with model with 9 inputs as an example. At the same time both models have absolutely identical learning results judging by the metrics. That is, both models come to the area of generality, only one from the bottom, the other from the top. And which one do you think will be the best? You will say the one that is simpler and has fewer inputs???? No. The one with more inputs will be cooler. Since they are both in the same found region, the one with more inputs will be more parametric with respect to the smaller one. As a result we get two models that are in the same state, just one strong, the one at the top, and one weak, the one at the bottom. But the results of their learning are equal. This means that the stronger is the model that will require more knowledge of the found law, and not the one that neglects the additional inputs. IMHO naturally!!!
 
Mihail Marchukajtes:
I'll tell you this. The last version of Reshetov was the 14th. I had enough courage to complete the optimizer to version 15. But what I added there changes the picture for the better. Proceeding from the theory that the most robust model is the one with minimal inputs and shorter polynomial, all other things being equal, I took a different path. If Reshetov's solution was model ascending, i.e. model ascending from 4 inputs and then adding more inputs so we obtain a model with say 7 inputs, I started with descending, namely, I start with 11 inputs and then decrease and finally I obtain a model with 9 inputs, say, as an example. At the same time both models have absolutely identical learning results judging by the metrics. That is, both models come to the area of generality, only one from the bottom, the other from the top. And which one do you think will be the best? You will say the one that is simpler and has fewer inputs???? No. The one with more inputs will be cooler. Since they are both in the same found region, the one with more inputs will be more parametric with respect to the smaller one. As a result we get two models that are in the same state, just one strong, the one at the top, and one weak, the one at the bottom. But the results of their learning are equal. This means that the stronger model that will require more knowledge of the found law, and not the one that neglects the additional inputs. IMHO naturally!!!

It can be different, you have to look at the statanalysis. The market has very few dimensions in the quotes themselves, 3-7 is normal.

 
All in all, I had a good day today. I think I made a full automatic after all. I got rid of the division to zero in the robot. The money is credited instantly, although they said that you have to wait three days. I had to wait three days for the money to come in. In a word, a fairy tale. I have been working with my indicator calls lately it was too depressing but I finally succeeded in solving them today. EEHHHHHHHHHHF FORTS hang on, Misha is coming :-)
 
Wow, you're so good, you're hilarious))))
 
Vizard_:
Wow, you're so good, you're hilarious)))
Wow, what a bunch of people. You know your smile warms our souls, especially when we haven't seen each other for so many years :-)
 
Mihail Marchukajtes:
Wow, what a bunch of people. You know, your smile warms our souls, especially when we haven't seen each other for so many years :-)

How can I see you if I quit my job and bought a camera, but still no videoclips, you're playing the keyboard the old-fashioned way)))
Come on videopilis and tell me about the latest successes and 15 super version!!!

 
Vizard_:

How to see you if you quit your job and bought a camera, but still no videos, all the old way - torturing the keyboard))))
Come on videopilis and tell me about the latest successes and 15 super version!!!

Yes it will, it will. Soon. When I started preparing for the lecture I accidentally wrote a book, but most importantly I want to familiarize you with the theory of retraining. To bring to public judgment so to speak :-)
 
Well, here's the first deal opened, so you can say me with a start :-)
  
void OnTick()
  {
// Получим ценовой прогноз от нейросети
   Prognosis=CalcNeuroNet();

// Осуществим необходимые торговые действия
   Trade();
  }
//+------------------------------------------------------------------+
void Trade()
  {

// Закроем открытую позицию, если она против прогноза

   if(PositionSelect(_Symbol))
     {
      long type=PositionGetInteger(POSITION_TYPE);
      bool close=false;
      if((type == POSITION_TYPE_BUY)  && (Prognosis <= MinPrognosis)) close = true;
      if((type == POSITION_TYPE_SELL) && (Prognosis >= -MinPrognosis)) close = true;
      if(close)
        {
         CTrade trade;
         trade.PositionClose(_Symbol);
        }
     }

// Если позиций нет, то откроем по прогнозу

   if((Prognosis!=0) && (!PositionSelect(_Symbol)))
     {
      CTrade trade;
      if(Prognosis >  MinPrognosis) trade.Buy (Lots);
      if(Prognosis < -MinPrognosis) trade.Sell(Lots);
     }
  }
Help me and write how to make the Expert Advisor open trades only on GBPJPY. And analyze the chart where it stands. Thank you !
1...160916101611161216131614161516161617161816191620162116221623...3629
New comment