Machine learning in trading: theory, models, practice and algo-trading - page 962

 
Maxim Dmitrievsky:

I just have prices at the entrance, chips do not suffer :) the main thing is the selection of targets

MaximDmitrievsky:

I do not know what to do next, no one throws up ideas, too lazy to think

ideas?

you feed prices, price deviations, logarithm of price.... and what should be the output? - imho, the maximum quote mechanism, that's what you can find on the small TF

try to guess what will be the first updated (broken) on the new bar High[1] or Low[1] - if you learn to predict, then it's a profit, because you already know the direction of price

;)

 
Igor Makanu:

ideas?

feed prices, price deviations, price logarithm.... and what should be the output? - imho, the maximum quoting mechanism, that's what you can find on the small TF with the price

try to guess what will be the first updated (broken) on the new bar High[1] or Low[1] - if you learn to predict, then it's a profit, because you will already know the direction of price

;)

Quoting mechanisms - that's a separate cheat direction :)

I would like to do it on timeframes of 5 to 15 min. I have already shown on the screenshots that it catches some regularities unrelated to quoting. But it works not very long on the OOS, at most 2X trace

It is interesting to look at the breakdown, I should try it.

 

History/Future = 30000/1000. Input - time series of differences Close and Open: iClose(NULL, 0, i+j+1+Shift) - iOpen(NULL, 0, i+j+1+Shift), where j is from 0 to 99 (Total 100 pieces). The target is the color of the bar (0,1).

On charts only OOS period (Future).

There is no spread. Level for entering an order - 0.


Spread - 2 points. Level to enter an order - 0.

Spread - 2 points. Level to enter order - 0.1

Spread - 2 points. Level to enter the order - 0.15.

Distribution of predicates between classes. Accuracy - 0.525.


 
Ilya Antipin:


Distribution of predicates between classes. The accuracy is 0.525.

Some sample is tiny - I have a 100-200 thousand sample, and if I take a slice of 1000, there's a good chance that there will be a better conjugacy.

 

It is very easy to make a genius thing. It is difficult to reach the state in which genius things are made. Amedeo Modigliani :)

That's why you have to try different kinds of herbs

 
Maxim Dmitrievsky:

Well there is a model, trained steadily well in different modifications, some at 100% and more from trayn work out on OOS, as here... (4 months of training 10 months of OOS) then shit

I do not see the point in testing the demos because everything is already clear.

Yes, it's a pity that your demo was lost. And all because you are on the CB too much looking, even though he brought the article, which says that the CB to choose the model can not, and on the forum a lot of times here wrote the same thing.

 
Dr. Trader:

Yes, it's a pity that your demo was lost. And all because you are too much looking at the OOS, although the very same article where it was written that the OOS model can not choose, and on the forum here many times wrote the same thing.

Here's what it turns out...

I will copy my reasoning/suggestions on this subject from another thread:

It seems to me a little too much of EP for evaluation of the model, that's why I wondered why you do the selection only on this part.
Yes, it works in a particular case (you got good results in all segments), but it seems to me it is not universal.

After all, you could come across data that are not as good. And the model could, for example, learn up to 40% of error in the training area, and purely by chance show 30% of error in the test area. And let's say the second model learned up to 35% in both cases. The second model is obviously better. But selecting only the test plot will select the first one. For comparison, there are options for evaluating the model:
evaluation only on the training plot,
or on the sum of all plots,
or as in Darch, (at submitted validation data) by Err = (ErrLeran * 0.37 + ErrValid * 0.63) - these coefficients are by default, but they can be changed.

The last option is the most interesting, because it takes into account both errors, but with a large weight of the validation section.
You may extend the formula, for example, to Err = (ErrLeran * 0.25 + ErrValid * 0.35 + ErrTest * 0.4).

Maybe we should even make a selection by delta errors, for example, if ErrLeran and ErrTest differ by more than 5% - we should reject such a model (San Sanich was talking about it). And from the rest to make a choice.
 
elibrarius:
Here's what it turns out...

I will copy my reasoning/suggestions on this topic from another thread:

It seems to me a little bit too much of the EPO to evaluate the model, that's why I wondered why you were doing the selection only on this section.
Yes, it works in a particular case (you got good results in all segments), but it seems to me it is not universal.

After all, you could come across data that are not as good. And the model could, for example, learn up to 40% of error in the training area, and purely by chance show 30% of error in the test area. And let's say, the second model could manage to learn up to 35% in both cases. The second model is obviously better. But selecting only the test plot will select the first one. For comparison, there are options for evaluating the model:
evaluation only on the training plot,
or on the sum of all plots,
or as in Darch, (at submitted validation data) by Err = (ErrLeran * 0.37 + ErrValid * 0.63) - these coefficients are by default, but they can be changed.

The last option is the most interesting, because it takes into account both errors, but with a large weight of the validation section.
In principle, you can extend the formula, for example to Err = (ErrLeran * 0.25 + ErrValid * 0.35 + ErrTest * 0.4).

Maybe we should even make a selection on delta errors, for example, if ErrLeran and ErrTest differ by more than 5% - then to reject such a model (San Sanich told about it). And from the rest already make a choice.
IMHO, in addition to formulas for summing up the errors, we need more proportions of their relations, it seems that someone already wrote here that the errors in the sections should correlate as Train <= Valid <= Test.
 
Ivan Negreshniy:
IMHO, in addition to formulas for summing up errors, we need more proportions of their ratio, here it seems someone already wrote that the errors by sections should correlate as Train <= Valid <= Test.
Train is minimized by the error of training, the other areas may randomly wobble up and down. The main thing is that it should not be too much.
 
elibrarius:
Train is minimized by training error, the rest of the sections can randomly dangle, either up or down. The main thing is not to dangle too much.

And what do you, for yourself, justify the permissibility of downward deviations, other than pure chance?

And then what is your main task, if not to combat this randomness, because it negates the meaning of both validation and OOS and MO in general.))