Machine learning in trading: theory, models, practice and algo-trading - page 3632

 
Maxim Dmitrievsky #:
Feature selection by brute force methods is such a kozul inference on minimax :) scientists are just now racking their brains over it. That is, they have come to the conclusion that it does not work well and they need to do something about it :) they even organise contests.

In the meantime, let's trust the omnipotent random :)

 
Aleksey Vyazmikin #:

There are no formulas that can reliably describe this process. I think wrong expectations can lead to wasted effort. You can think of a price chart as a sound wave from a rubbish recycling machine, and you can learn to tell by the sound what the machine is shredding now, but what is next in line is not so easy without data beyond those sound waves.

Here on the forum, the term "overtraining" is often misused - and implies just a bad result. In reality, though, the term implies a strong fit to the training data, licking all points with great precision. For trees, depth splits up to 100% of the class representative in the leaf. And in fact, the problem is often a change in the probability distribution, i.e. it is the observed data from which the predictor was calculated. I showed everything in detail in this thread at the beginning of the year.

Of course, "there are no formulas". Then what is there when you think the network has successfully worked out the new data? It's just an approximation and nothing more. But an approximation to an approximation to an approximation to a wolf, you know yourself.... I think you know about the theorem, too.

But, tell me about the key, yellow metal is always valuable.

Why should I tell you that? I've shown the contradictions that classical MO methods can't solve. You're searching, but don't think there's a magic MO method that will solve everything, you just need to read the right book. Such methods do not exist, for obvious reasons, especially in the case of TCDC.

Nonstationarity is such a strange thing, so do you want me to give you a series, just for fun, and you try to predict the next steps of the series? In this case I will show that the formula is the same, but you will not be able to predict the series by your MO methods. Not because MO methods "don't work", but because you refuse to accept that even a strict simple formula is hard to predict if you use the wrong evaluation criteria.

Even if you do everything right, buy the right cool car, fill the car with the coolest petrol, you will still arrive at the wrong place, all because the targeting was wrong. No matter what you do, you will come to Peter, not to Ekaterinburg, if your navigator in the automatic driving system is set incorrectly. The targets are set wrong, it's strange that everyone is surprised that they don't arrive where they want to go.

 

The picture, which has become a textbook and explains why financial VRs are non-stationary at any TF and at any number of observations, i.e. the estimation does not depend on whether we know the whole history (future) or not.

This is the basics

Just like an English-speaking moderator cannot know the context of posts, but deletes them on constant petty complaints of storytellers )))

Then let him communicate with himself here himself

 
Maxim Dmitrievsky #:

The picture, which has become a textbook and explains why financial VRs are non-stationary at any TF and at any number of observations, i.e. the estimation does not depend on whether we know the whole history (future) or not.

Just as an English-speaking moderator cannot know the context of posts, but deletes them on constant petty complaints ))

and now take the square root of the picture and wonder that returns are the equivalent of activity (I don't know what else to call it, market activity), except for the linear momentum in the peaks and the inability to drop below the limits when "nobody is home" :-)

just don't take the "average temperature of the hospital during the year", you should consider individual known and understandable states.

 
Maxim Kuznetsov #:

and now take the square root of the picture and wonder that returns are the equivalent of activity (I don't know what else to call it, market activity), except for the linear momentum in the peaks and the inability to drop below the limits when "nobody is home" :-)

just don't take the "average temperature of the hospital during the year", you should consider individual known and understandable states.

That's different )

then we need to add systematic uncertainty, when stationarity seems to be found, but due to changing conditions does not remain so on new data.
 

I asked if MO experts would be willing to train a grid on a process that has a strict formula, and then validate their model on new data (just a time offset t on the same formula)? It's just a formula, pure as a child's tear. And, as a bonus to self-education, you can measure any distributions of that process and compare what it was Before and what it was After.

I hope for a positive response from those willing to participate in the experiment. Otherwise, I credit the drain to the "experts" in MO. I emphasise that the formula of the process will be simple.

 
Aleksey Vyazmikin #:

In the meantime, let's trust in the almighty random :)

Seems like yesterday/today there should be an award in the contest. Went in - haven't seen any results yet.

Interesting to read the winners/me comments afterwards.

If they actually showed a good scor on that data, that's strong.
 
Andrey Dik #:

Of course there are "no formulas." Then what is there when you think the network has successfully worked out the new data? It's just an approximation and nothing more. But an approximation is an approximation of an approximation of a wolf, you know..... About the theorem, too, I think you know.

I think that a similar tin has been caught, and the model knows how long to grind it and how it will rattle.

Personally, I don't use nets now, and I don't predict the bar closing price.

Andrey Dik #:
do you want me to give you a series, just for fun, and you try to predict the next steps of the series?

You did not succeed?

Andrey Dik #:
a strict simple formula is difficult to predict if you use the wrong evaluation criteria.

How do you select a different evaluation criterion for each formula?

 
Maxim Dmitrievsky #:

Seems like there should be an award in the contest yesterday/today. Went in - haven't seen any results yet.

Interesting to read comments from the winners/me afterwards.

If they actually showed a good scor on that data, that's strong.

I didn't follow the contest. I think such contests are a kind of brainstorming, the organisers will get food for thought anyway.

 
Aleksey Vyazmikin #:

1. I think a similar tin has been caught, and the model knows how long to grind it and how it will rattle.

2. Personally, I don't use nets now, and I don't predict the bar closing price.

3. You have not been able to do this?

4. How do you match each formula to a different criterion for evaluation?

1. It is easy to check.

2. Zakr bar not prchm. At all. In n and ns ses. But you understood.

3. What do you think if I asked you what made you think?

4. That's not required. You don't have to pick anything up.