Machine learning in trading: theory, models, practice and algo-trading - page 3649

New comment
 
Maxim Dmitrievsky #:

There is such a question on variable length traits.

Would this be any different from splitting BP into multiple states?

After all, in essence, different lengths of attributes should somehow differentiate these states for the model.

From this point of view, there is no point in bothering, is there? If there is already a division into states.

What does it mean - different feature lengths?
 
Ivan Butko #:

Maybe what is meant is that if you retrain for 20 years, the first 10 years have the same set. It turns out that the second 10 years is an extrapolation if the set is found in some way that does not include the second period (about retraining - as an example of the set working on a large period).

Ivan, you know this yourself, but I'll write it for those who don't know.

This is what approximation and extrapolation imply - approximation to a process whose nature is strictly unknown. There are many other ways than neurons to describe a process or its properties with approximation, such as random forest or boost. Ultimately all methods are designed to approximate a process, including statistical methods. The goal is to get information about the process and use the information on new data - extrapolating the process itself or its properties to new data. This is what MO does by definition.

 
Maxim Dmitrievsky #:

Neural networks can't extrapolate at all.

Why not? Even boosts are not only neuronics can do it
 
mytarmailS #:
What does it mean to have different feature lengths?
When the number of features changes over time, the number of columns. That makes more sense :)
 
mytarmailS #:
Why is that? Even boosts can do more than just neural nets.
Well, try extrapolating a series of natural numbers with a neural net ) boosts are even worse at it.
 
Ivan Butko #:

Maybe what is meant is that if you retrain for 20 years, the first 10 years have the same set. It turns out that the second 10 years is an extrapolation if the set is found in some way that does not include the second period (about retraining - as an example of the set working on a large period).

You came closest, but it's called interpolation. NSs are good at interpolating and bad at extrapolating. A function is interpolated.
 
mytarmailS #:
Why is that? Even boosts can do more than just neurons

I gave Dick a link to a simple tree training code. If he had spent 20 minutes studying it, he wouldn't have spent 4 days writing here his fantasies about extrapolation to MO.
Well, he obviously came here to troll....
But it turns out that the regulars of this thread have not seen the code.

I'll explain in a very simple way. There was an example of teaching a forest the multiplication table from 1 to 9. Using his example:

If you trained up to 9*9 at most, then the corresponding leaf will always answer 81. And at queries 10*9 and 9*10 and 11*20 and even 1000*2000 the answer will always be 81. There are no leaves with answers 90,100,220, 2000000.
What extrapolation? The answers will only be from 1 to 81.
 
Forester #:

I gave Dick a link to a simple tree training code. If he had spent 20 minutes to study it, he would not have written here for 4 days his fantasies about extrapolation to MO.
Well, he obviously came here to troll....
But it turns out that the regulars of this thread have not seen the code.

I'll explain in a very simple way. There was an example of teaching a forest the multiplication table from 1 to 9. Using his example:

If you trained up to 9*9 at most, then the corresponding leaf will always answer 81. And at queries 10*9 and 9*10 and 11*20 and even 1000*2000 the answer will always be 81. There are no leaves with answers 90,100,220, 2000000.
What extrapolation? The answers will only be from 1 to 81.
+
 
Using the example of an integer multiplication table from 1 to 9 - you approximate the f- function with a neural network and it starts solving examples with fractional values because they are all inside the function. Like 1.2 * 2.4. But if you start multiplying by 10, 20, you won't get anything.
 
I think people confuse the generalisation ability of NS with extrapolation. Extrapolation implies going beyond the boundaries of the training set and as a consequence either poor results or no results at all (depends on the type of NS). Generalisation implies good performance of the network on new data, but which do not go beyond the boundaries of the training set.
Therefore, networks approximate the data on the training sample and interpolate on new data that do not go beyond the training dataset.
1...364236433644364536463647364836493650365136523653365436553656...3700
New comment