Machine learning in trading: theory, models, practice and algo-trading - page 2626

[mytarmailS]

Maxim Dmitrievsky #:
Just show me if you got it, I'll show you my results on the method, I had no time to finish it.

It's still in development, a very expensive process in terms of computer resources, while I'm optimizing the code

[mytarmailS]

Interesting article, something to think about

https://www.autodesk.com/research/publications/same-stats-different-graphs

[mytarmailS]

The importance of signs in the moving window (indicators and prices)

At one moment the indicator may be 10% important and at another moment it may be 0,05% important, such is the truth of life)

If you think it solves everything, you should be proud of it.

That's what the four signs of Fisher's Iris look like.

Or if you enlarge the sliding window like this

[Forester]

mytarmailS #:

The importance of signs in the moving window (indicators and prices)

At one moment the indicator may be 10% important and at another moment it may be 0,05% important, such is the truth of life)

If you think it solves everything, you should be proud of it.

That's what the four signs of Fisher's Iris look like.

Or if you zoom in on the sliding window.

It is clear that the irises (and similar problems) have a stable pattern. Everyone who's been experimenting with them has already figured out that everything "floats" in quotes.

I wonder how the significance of the indicators is different in every point of the chart. It is determined for the whole model built on all training lines at once. Or do you have 5000 models there?
And in general, explain your graphs, what is on them and how they were built.

[mytarmailS]

elibrarius #:

The fact that irises (and similar problems) have a stable pattern is already clear. And the fact that everything "floats" in the quotes is also clear to everyone who has experimented with them.

I wonder how the significance of the indicators is different in every point of the chart. It is determined for the whole model built on all training lines at once. Or do you have 5000 models there?
And in general, explain your graphs, what's on them and how they were built.

There are a lot of ways to find out the informativeness of a feature, some you don't have to train a model for. I used fselector. https://www.r-bloggers.com/2016/06/venn-diagram-comparison-of-boruta-fselectorrcpp-and-glmnet-algorithms/

It counts the entropy of the features...

Why is the importance different at each point? Because the informativeness of the features was calculated in a sliding window as I wrote above.

[Maxim Dmitrievsky]

mytarmailS #:
There are a lot of ways to find out trait informativeness, for some you don't have to train a model. I used fselector. h ttps://www.r-bloggers.com/2016/06/venn-diagram-comparison-of-boruta-fselectorrcpp-and-glmnet-algorithms/

He counts the entropy of features...

Why is the importance different at each point? Yes, because the informativeness of the features was calculated in a sliding window as I wrote above.

So you have to look for periods where the importance does not jump, you can use 2 models. Otherwise, it's a mess.

I did online window training, if you take it all together without filtering by time, performance is poor. I didn't think of doing it with filtering at the time. There is an example of such bot in my article about entropy

Most likely, importance jumps are due to entropy changes, if signs like returns

But all the adepts of foregate gates have their own reality not based on practice

[mytarmailS]

Maxim Dmitrievsky #:

But all sorts of adepts of foregate have their own reality, not based on practice

What is that?

I think it is necessary to look for a pattern and build a model for it, often MO cannot build a model even for an understandable pattern, all by hand

[Maxim Dmitrievsky]

mytarmailS #:
What is that?

I think you need to look for a pattern and build a model for it, often the MO cannot build a model even for an understandable pattern, you have to do it all by hand

Well, there are all sorts of recurrence networks, there was one here

straight through the pattern and look for a pattern where it behaves in a pattern :)

Quite simply: train it, check it on a test, identify periods where it was shedding and working, draw conclusions/try to filter it out, identify a pattern

You should not separate statistics from the MO, you should use statistics for models, they are trained randomly

If you know the pattern, you kind of don't need the MO

[mytarmailS]

Maxim Dmitrievsky #:
You should go straight to the model and look for a pattern, where it behaves as if it were a pattern :)

If very simple: teach it, test it on the test, identify periods where it was pouring and working, draw conclusions / try to filter it out, identify a pattern

Yes, in principle it is possible, even better, in this order you can do on the machine

Maxim Dmitrievsky #:

If quite simple: teach, test, identify periods when it was pouring and working, draw conclusions / try to filter out, detect a pattern

or not to pour))

Maxim Dmitrievsky #:

You should not separate statistics from the MO, you should use statistics for models, they are trained randomly

For me, it is not necessary to make complicated models, a simple rule is enough, otherwise you cannot call it a pattern.

Maxim Dmitrievsky #:

SZY if you know the pattern, then the MO is kind of unnecessary.

I always want to do better)))

[Forester]

mytarmailS #:
There are a lot of ways to find out the informativeness of a feature, some you don't have to train a model for. I used fselector. h ttps://www.r-bloggers.com/2016/06/venn-diagram-comparison-of-boruta-fselectorrcpp-and-glmnet-algorithms/

He counts the entropy of features...

Why is the importance different at each point? Yes, because the informativeness of the features was counted in the sliding window as I wrote above

I was comparing several ways to estimate the importance of the attributes. I took the most resource-intensive one as a benchmark: training the model by removing the features one by one.
The fast methods do not coincide with the benchmark. They do not match each other either. The fselector is even faster, I don't think it will match anything either.