Machine learning in trading: theory, models, practice and algo-trading - page 2010

 
Aleksey Vyazmikin:
Can we hold a contest for sampling training?

Let the market judge.... you couldn't have said it better

If you open a signal, keep 3-4 months in the black with reasonable drawdowns, you will have a contest and competition, and later you will get a reward in the form of subscribers

imho, even a demo account is enough if the initial deposit is reasonable ;)

 
Valeriy Yastremskiy:
Samples need to be formalized somehow. Otherwise, suddenly the phasers will be involved and win).

I think that no OHLCVs should be given, and new predictors cannot be added, only combinations, merges, and selections can be made from existing ones. The purpose of the contest would be to apply learning skills with disclosure of learning on available data.

 
Igor Makanu:

let the market judge.... there is no better way to tell

Open a signal, keep 3-4 months in the black with reasonable drawdowns, you will have a contest and competition, and later remuneration in the form of subscribers

imho, even a demo account is enough if the initial deposit is reasonable ;)

I am interested not to assert myself, but to broaden my horizons, and in general to compare different approaches of training to one sample.

Yes, and if I am an organizer, then I can give my sample, and the results for me will be doubly interesting.

 
Aleksey Vyazmikin:

I think no OHLCVs should be given, and no new predictors can be added, only combinations, merges, and selections can be made from existing ones. The purpose of the contest will be to apply learning skills with disclosure of learning on available data.

One sample for all norms. How to implement with future sampling. Good, only days or weeks of training in the real world and then the real world. 4 hours may not be enough for training. Well and the same reasonable depth of sample for training.
 
Valeriy Yastremskiy:
One sample for all normals. How to implement with future sampling. In a good way in the real only on the days or weeks of training and then on the real. 4 hours may not be enough time for training. Well and the same reasonable depth of sample for training.

I think the sample would be in the neighborhood of 10,000 rows, and 5,000 to check the results of the training - production, which will be provided after the time for training is over.

The question is how to fix the models, because everyone will teach using different methods, who on what.

 
Aleksey Vyazmikin:

I think the sample will be in the neighborhood of 10,000 lines, and 5000 to check the results of training - production, which will be provided after the end of time for training.

The question is how to fix the models, because everyone will teach different methods, who uses what.

The publicity of the code unconditional victory. Not publicity further public testing on the real.
 
Aleksey Vyazmikin:
Can we hold a contest for sampling training?

This reminds me of someone else.)

Igor Makanu:

let the market judge.... you couldn't have said it better

open a signal, wait 3-4 months at ....

Why wait months if you can check it right away?

Aleksey Vyazmikin:

I think the sample will be in the neighborhood of 10,000 lines, and 5,000 to check the results of the training - production, which will be provided after the end of time for training.

The question is how to fix the models, because everyone will teach by different methods, who on what.

By acurasi to fix, what does it matter what model and what it is written on?

Aleksey Vyazmikin:

I think that no OHLCVs should be given, and no new predictors can be added, only combinations, merges, selections can be made from existing ones. The purpose of the contest would be to apply learning skills with disclosure of learning on existing data.

nonsense! if everyone uses the same data with no possibility to improve anything, then the difference in the results will be measured in tenths of percent, regardless of the models, there is no point in such activities

Such combinations are made automatically by the algorithm, it makes no sense to do them.

 
Valeriy Yastremskiy:
Publicity code unconditional victory. Not publicity further public testing in real life.

Publicity of the code of training is unnecessary - the description of the approach is more important to me. Simply the model itself will have to be fixed and posted publicly.

 
mytarmailS:

Someone reminds me of this )

Acuracy, what difference does it make what model and what it is written on?

Bullshit! If everyone is going to lick the same data without being able to improve it, the difference in the results will be measured in tenths of a percent, regardless of the model, there is no point in such activities

Such combinations are made automatically by the algorithm, it makes no sense to do them

Were there any such contests in this thread?

I'd like to use a sample with 3 target values, so accurasi won't be enough.

I'm just now interested in methods of post-processing the data obtained - this information can be shared. Or are you willing to share information about your predictors? The result you get can be quite different.

I was talking about methods to reduce dimensionality, to allocate predictor ranges, to search for patterns through clustering, to increase sampling by past data, to use prediction results on past data, and other methods that give new information from the existing one.


Oh, and OHLCV won't be there for the reason that the sample will be built from 2014 to 2020 on minutes.

 
mytarmailS:

The cuts began to coincide +- with the tester ... well, at least, if the real is not good, then in the tester too

For now it is so since September 21, in the tester it is about the same. And for an earlier period in the tester it is good, but the weeks are unsuccessful (probably)

For some reason in the beginning of the week it is always good (after the "pre-training"), then it "drops", maybe just coincidence

found something else to improve, maybe it will finally work )

long fucking wait to catch bugs

on the long movements it pulls, on the U-turns stalls

here is an example of the tester (updated logic) the last 2 months

But it is worse with old logics. I will transfer the new logic to the trader and lock it next week, maybe.