Machine learning in trading: theory, models, practice and algo-trading - page 1546

 
Maxim Dmitrievsky:

overlearning in the sense of weak generalization. Above already wrote how you can get around the problem, but there are more elegant approaches, I'm sure

With the very quality of training on the train + valide no problem at all

So maybe it's about the data, it's not the first time I hear from different trainers that homogeneous data, such as increment, better feed the NS, and trees of different types work better with non-uniform data - patterns, news, risk factors, time, events, stock density, open interest, volumes.

If you want to handle this, you'll need to implement an analyzer with the same methods as in the previous one.

 
Sergey Chalyshev:

I see everyone is trying to train the network with the help of a teacher.

Has anyone tried to train on a target function, such as the recovery factor?

I select leaves and build a model from them according to my parameters.

 
elibrarius:

It doesn't suggest it.

In XGBoost, the first tree is the rough model. The rest correct the first one, and with a microscopic factor. You can't get anything working separately there, they only give good results as a whole.
In Katbust apparently the same basic principle, with its own features.

In fact, and I am skeptical about it, except to make the tree more authentic - now I prepare the data for 6 splits, I think it is not enough.

However, the very essence of the weight is just the evaluation of all the sheets in the model on an accrual basis, and you can not exclude that there is a good pattern among them, because the principle of building sheets is observed and takes into account the independent construction of greed, and then check for tree improvement and its evaluation. Let's see.

 
Aleksey Vyazmikin:

So maybe it's all about data, it's not the first time I hear from different lecturers that homogeneous data, such as increments, is better to feed NS, and trees of different types work better with heterogeneous data - patterns, news, risk ratios, time, events, the cup density, open interest, volumes.

By the way, about the increments, have you tried to measure not in pips, but by ATR, or as a percentage of closing price?

I don't think so... I don't care what you measure.

 
Maxim Dmitrievsky:

You're struggling with the wrong thing... I don't give a shit what to measure in.

Just the opposite, I thought converting to natural values would have an effect, because I have all values normalized and quantized (broken down into ranges), and it turns out that when I left the pure numbers, the learning deteriorated significantly. It's obvious to me now, that preprocessing of data makes a difference.

 
Aleksey Vyazmikin:

Just as I thought that on the contrary, conversion to natural values would have an effect, because I have all values normalized and quantized (broken down into ranges), and it turned out that when I left the pure numbers, the learning deteriorated significantly. It's obvious to me now that preprocessing the data makes a difference.

Well, you have your own bizarre world there, with its own beasts )) I only use increments and their counterparts, and sometimes just prices, as the Fathers commanded

 
Maxim Dmitrievsky:

Well, you have your own bizarre world there, with its own beasts )) I use only increments and their analogs, and sometimes just prices, as the Fathers commanded

Maybe I will cross two samples with your and my predictors, just for the sake of experiment?

 
Aleksey Vyazmikin:

How about crossing two samples with your and my predictors, just for the sake of experimentation?

Why? Any predictors are derived from returns. Just add returns to yours and consider them crossed already

 
Maxim Dmitrievsky:

Why? Any perdictors are derived from returns. Just add returns to yours and consider it crossed already

I don't know what returns to add, at what pitch and how many pieces.

 
Aleksey Vyazmikin:

I don't know what returns to add, at what pitch and how many pieces.

I do not know, always different