Retraining - page 7

 
Youri Tarshecki:
Yousufkhodja Sultonov:

Friends, no need to argue.
The purse will decide who is right :)
 
Event:
The purse will decide who's right :)
Yeah.)
 
Комбинатор:
So yeah )

No.

It is much easier to test the performance of an EA on non-optimised sections of history than to waste money on tweaking EAs.

By the way, our character has never shown a successful wolfing forward of his EA that he suggests to buy for $1500 and wait for decades for profit.

 
Youri Tarshecki:

No, we start with, say, 1975.

Optimisation 1975-1985, verification 1985-1990

Optimisation 1980-1990, verification 1990-1995

Optimisation 1985-1995, check 1995-2000

Optimization 2000-2005, verification 2005-2010

Optimization 2005-2010, verification 2010-2015

Look only at test results, and if at least ONE of these five years will be negative (and I think there will be more), then the system is defective.

I.e. your trick with fitting on the whole history will only work if there are crazy people ready to wait for decades for profits from your EA.

And by the way, don't forget to tell us how you avoid over-optimisation on each plot.)

You've been going on for decades, you're going to live forever.
 

I will join your discussion, especially because the topic is very topical and interesting. And then if we re-train them under the same conditions we will get a different model that will behave exactly the same as the previous one, but on future quotes of these two seemingly identical models will work differently. The question is how to choose the model that will work in the future. A net must correspond to the market in the training area. And that grid, which has a larger prognostic variable, is more adequate to the current market situation. My NS classifies signals from the TS. There are about 10 signals per day but to choose which model to use I do the following. I consider the prognostic variable of the network operation in the optimization area and the value that is big for the model, that model is used.

Suppose the model value is up, i.e. current value is higher than the previous one, and the NEXT bar is also up. I.e. if the grid has predicted growth, we add one to the variable, if not, we subtract it and apply the same procedure for going down. It means we look up the prognostic variable of our model and which model has higher number, it means the model more often predicted the market, so it describes it better and we choose..... in the code looks like this

double PONT11=iCustom(NULL, 0, "Модель",1,i)-iCustom(NULL, 0, "Модель",1,i+1);
if ((PONT11>0)&& (Close[i-1]>Open[i-1])) AA=AA+1;
if ((PONT11>0)&& (Close[i-1]<Open[i-1])) AA=AA-1;
if ((PONT11<0)&& (Close[i-1]<Open[i-1])) AA=AA+1;
if ((PONT11<0)&& (Close[i-1]>Open[i-1])) AA=AA-1;

So... that's it... anyone has any thoughts on this. I'd like to hear an opinion....

 
Youri Tarshecki:

No, we start with, say, 1975.

Optimisation 1975-1985, verification 1985-1990

Optimisation 1980-1990, verification 1990-1995

Optimisation 1985-1995, check 1995-2000

Optimization 2000-2005, verification 2005-2010

Optimization 2005-2010, verification 2010-2015

Look only at test results, and if at least ONE of these five years will be negative (and I think there will be more), then the system is defective.

I.e. your trick with fitting on the whole history will only work if there are crazy people willing to wait for decades for profits from your EA.

And by the way, don't forget to tell us how you avoid overoptimization at each area.)

Optimisation 1975 -1985 (optimal sample size = 80 bars of history):

1985-1990 check:

Optimisation 1980-1990 (optimal sample size = 80 bars of history):

1990-1995 check:

Optimisation 1985-1995 (optimal sample size = 360 bars of history):

1995-2000 check:

Optimisation 2000-2005 (optimal sample size = 330 bars of history):

2005-2010 verification:

Optimisation 2005-2010 (optimal sample size = 330 bars of history):

2010-2015 verification:

Didn't meet your expectations, all verification sites are overcome with positive results, although not outstanding.