Machine learning in trading: theory, models, practice and algo-trading - page 3098

New comment
 
Maxim Dmitrievsky #:
Let it be known that MO ts are no different from the rest, the success rate is on average the same (about zero, but sometimes ploughs on)

But after MO, to switch back to the indicator ones, it's like going from a Mercedes to a Zaporozhets. It seems to drive, but the sensations are not the same.

Bonus is brain training, if you don't want to suffer from marasmus in old age and keep potency :)

You become (well, not you specifically) really smart and look at the world differently. So even if it doesn't work out, you still get dividends.

I agree with the highlighted part.

not everyone has that:

No, not Neira (her processor would rather boil than find such a solution), I did it myself.

 
Maxim Dmitrievsky #:


Due to the fact that we remove the bias in the training sample (this is the main thing) and variance through cross-validation, the model starts to behave +- adequately on new data. Then it can be fine-tuned.


by the way, have you tried to make a chart not with a uniform step between trades, but by time?
Or it may turn out like mine with 5 years of only 2 areas of growth for half a year, the rest of the time almost without trades. And 2 years in drawdown, for the same reason. You can't put such a thing on the real...

If you do it not by time, but by steps, it will be as beautiful as yours.

 
Forester #:

by the way, have you tried to make a chart not with an even step between trades, but by time?
Or it may turn out as I had for 5 years with only 2 growth areas for half a year, the rest of the time almost without trades. And 2 years in drawdown, for the same reason. You can't put such a thing on the real...

If you do it not by time, but by steps, it will be as beautiful as yours.

Your chips are probably out of range. Eurodoll chart below, it trades more or less evenly. But the OOS always has fewer trades, with equal length of training and OOS. Well, because the metrics are worse. I haven't managed to make it perfect yet.
 
Maxim Dmitrievsky #:
You've got the chips out of range there, I guess. From below eurodoll chart, more or less evenly traded. But the OOS always has fewer trades, with equal length of training and OOS. Well, because the metrics are worse. We haven't managed to make it perfect yet.

Try a time chart. It is not excluded that it will be the same....

 
Maxim Dmitrievsky #:

Kozul uses some external parameter to estimate its effect on the model results. It can be a predictor or a binary variable, anything. It can even be the difference of predictors.

Then, using different techniques, inference is made as to the effect of that parameter on the predictions. After that, one can uplift the model with this influence in mind, to get new values of, for example, labels. And new coefficients of the model, as it is done in double machine learning. There are 2 models there: one does debias, the other denoise. Since cross-validation is used in the estimation process, the new parameters are more robust, also on new data. Then the final model is trained.

It is difficult to explain on fingers, it is better to read special literature. I have made several variants, they work. The topic is quite large, with its own nuances. There are your favourite "packages".

There are both purely empirical approaches and strictly proven ones, like Chernozhukov's. It's a beautiful technique in general.



Due to the fact that we remove the bias in the training sample (this is the main thing) and variance through cross-validation, the model starts to behave +- adequately on new data. Then it can be fine-tuned.


There are plenty of different methods out there. Few prufs that they work. So the question is, now you are trying to find the areas of predictors, on which the variance increases in the aggregate and build a model to exclude them, and then on the residuals after classification you are training to apply the model in trade?

 
Forester #:

You try the timeline. It is possible that it will be the same....

There are no such large windows, the average is uniform, I checked. It happened on other TS when the signs went beyond the ranges.
 
Maxim Dmitrievsky #:
No such windows are large, average is even, checked. It was like that on other TCs when the signs went beyond the ranges.
I don't have them. The threshold is just high and cuts off a lot of trades. If you activate it through the middle, it will be much worse. In general, you yourself also raise the trading indicators (you wrote today).
 
Aleksey Vyazmikin #:

There's a lot of methods out there. There are few prufs that they work. So the question is, now you are trying to find the areas of predictors, on which the variance increases in the aggregate and build a model to exclude them, and then on the residuals after classification you are trained to apply the model in trade?

Can we just discuss kozul, not my ts

First the hard (for Sanych unbearable) stage of acceptance, then lapping, then love. 😀

 
Forester #:
I'm not getting out. Just threshold is high and it cuts off a lot of deals. If you activate it through the middle, it will be much worse. In general, you yourself also raise trade indicators (you wrote it today).
There was a lot calculation from the signal level.

 
Renat Akhtyamov #:

agree with the highlighted

not everyone has one:

No, not neura (her processor would rather boil than find such a solution), I did it myself.

Malevich, what the hell.
1...309130923093309430953096309730983099310031013102310331043105...3547
New comment
Reason: