Machine learning in trading: theory, models, practice and algo-trading - page 3574

New comment
 
Aleksey Vyazmikin #:

In response to my message that I had previously perceived NLP in one sense only and did not correctly understand your reference to NLP, you suggested that I was confusing the concepts, so I did not understand this suggestion. I assumed that you were talking about NLP (neuro-linguistic programming ) before, so I hinted that I am not a telepath and cannot understand what you were talking about....

Then when? If it was about your way of communication, something like that, then the first option, of course.
 
Maxim Dmitrievsky #:
When was that? If your way of communication was something like that, then the first option of course.

So I understood everything correctly and reacted correctly :)

 
A fresh and kind of interesting article on the topic of MO and non-stationarity.
 
Aleksey Nikolayev #:
A fresh and kind of interesting article on MO and non-stationarity.
Well basically it comes down to choosing a market mode and adding macro indicators to the chips.

It takes a very hardworking person to go through the main macroeconomic indicators and find the alpha in them :) And at a high level of awareness at that :)
 

There is an interesting article about the power of bousting. But I haven't found the library of the same name yet.

https://arxiv.org/abs/2407.02279

How to Boost Any Loss Function
How to Boost Any Loss Function
  • arxiv.org
Boosting is a highly successful ML-born optimization setting in which one is required to computationally efficiently learn arbitrarily good models based on the access to a weak learner oracle, providing classifiers performing at least slightly differently from random guessing. A key difference with gradient-based optimization is that boosting's original model does not requires access to first order information about a loss, yet the decades long history of boosting has quickly evolved it into a first order optimization setting -- sometimes even wrongfully \textit{defining} it as such. Owing to recent progress extending gradient-based optimization to use only a loss' zeroth ($0^{th}$) order information to learn, this begs the question: what loss functions can be efficiently optimized with boosting and what is the information really needed for boosting to meet the \textit{original} boosting blueprint's requirements? We provide a constructive formal answer essentially showing that...
 
Maxim Dmitrievsky #:
Well, it basically comes down to choosing a market mode and adding macro indicators to the chips. The first one is definitely useful, with the second one you should experiment, because they do not change as often as you would like. For example, once a month statistics is released. Then it will be a different, news-based TS.

You need a very hard-working person who would go through the main macroeconomic indicators and find the alpha in them :) And at the same time at a high level of awareness :)

I was attracted by the very approach, when one tries to work with non-stationarity in a meaningful way, studying its structure.

Obviously, when studying non-stationarity in a meaningful way, one can't get far on technical features alone (based only on prices). But meaningful and useful addition of non-technical signs is not an easy matter.

 
Maxim Dmitrievsky #:

There is an interesting article about the power of bousting. But I haven't found the library of the same name yet.

https://arxiv.org/abs/2407.02279

I should read it. I'm not sure that you can boost any loss function at all, but they may have come up with something interesting.
 
Aleksey Nikolayev #:

I was attracted by the approach itself, when one tries to work with nonstationarity in a meaningful way, studying its structure.

Obviously, when studying nonstationarity in a meaningful way, one cannot get far on technical features alone (constructed only by prices). But meaningful and useful addition of non-technical signs is not an easy matter.

Yes, there were such thoughts, but they went far ahead of hand.
 
Aleksey Nikolayev #:
A fresh and kind of interesting article on the topic of MO and non-stationarity.

Interesting but useless article. And the problem is that the approach to MO as a usual approach to statistics does not work.

I'll explain it with an example.

A few years ago I came across an article on garch, in which the choice of garch type (and there are more than 100 of them!) was made on prediction of all stocks included in S&P500. Everything is clear, statistics, you can use the results of the article and apply IGARCH more reasonably.

This does not work in MO.

It is not clear how the article tries to take into account some deviations, i.e. outliers.

At the level of preprocessing, the MO science suggests to somehow handle outliers, but the meaning of MO is lost behind these words: what are we teaching the model?

A seemingly simple question and a primitive answer.

However, it is far from it.

Some time ago, I brought to a demo account a TS with the teacher of price increment with classification error below 20%, or even 10%.

On the demo account I found out that this classification error corresponds to the ratio of losing and profitable trades.

It would seem that the profit factor should be about 4. But, no. Due to tricks in the Expert Advisor, I could not raise the profit factor above 1.5!

And here it became clear that my simple teacher taught the model to predict a profit of 10-15 pips, but was wrong on price movements over 50 pips.

For the last 1.5 years, I have been trying to construct a teacher (target variable) that would predict a profit multiple of a loss. But it turned out to be a very difficult task. I have to do a very intricate preprocessing of the teacher, then select a predictor for it, and it is extremely difficult to find predictors that are related to the teacher and not just noise..... All this in an overtraining and looking ahead environment - i.e., evaluation only on an external file with a step-by-step run...

Nothing like this in the article...

 

Sanych, a teacher is a person, not a target variable.

and you also deliberately emphasise your illiteracy by specifying"construct a teacher (target variable)".

1...356735683569357035713572357335743575357635773578357935803581...3590
New comment