Machine learning in trading: theory, models, practice and algo-trading - page 2651

New comment
 

I don't know about you, but I'm fascinated by this algorithm... How simple variables at the bottom add up to big complex expressions at the top, a hierarchy, just like in nature....

 
Aleksey Nikolayev #:

Yeah, the list thing is a good idea.

To make a real double, you need algebraic data types. I haven't seen their implementation for R yet, and genetic optimisation won't work with them( it's not clear what kind of optimisation will work with them atall).

Well, yes, there are limitations, because each digit for GP is a separate element in the list, so if there are a lot of them, then trouble(( , but you can get out of the situation as I have shown

 
mytarmailS #:

I don't know about you, but I'm fascinated by this algorithm... How simple variables at the bottom add up to big complex expressions at the top, a hierarchy, just like in nature....

How is it better than MSUA? You can be fascinated by anything, as long as it's useful.)
 
Maxim Dmitrievsky #:
How is it better than MSUA? You can be fascinated by anything, as long as it's good for you.)
Well, it's different.
Symbolic regression is a framework in which you can implement any Logic, it's just convenient because you don't have to write it yourself....

And this implementation of rules, yes it is very similar to mgua - feature ingeniring +feature selection+make model
All in one like mgua.
Well, if you compare this particular implementation with mgua:
MGUA can't handle a large number of features, as it's a complete overkill.
This is bad, but mgua finds a guaranteed best solution, which is good.

In this implementation it is exactly the opposite
 
mytarmailS #:
Well, it's different.
Symbolic regression is a framework in which you can implement any Logic, it's just convenient because you don't have to write it yourself....

And this implementation of rules, yes it is very similar to mgua - feature ingeniring +feature selection+make model.
All in one like mgua.
Well, if you compare this particular implementation with mgua:
MGUA can't handle a large number of features, as it's a complete overkill.
This is bad, but mgua finds a guaranteed best solution, which is good.

In this implementation it's exactly the opposite
Oh yes, it's genetic overkill and there are probably standard schemes for combining traits.
 

A little bit about invariance to events

Suppose we want to describe a head and shoulders pattern (let's assume that such a pattern exists) , our algorithm looks at the last H-points, let's say 7 as on the example

The green rectangle shows the scope of the algorithm. The last H-points, in this example 7 pieces.

The data in this form we process, feed into AMO, etc., sliding window, etc. Everyone knows it...

If we describe the regularity by rules, it will be something like x[1] < x[2] & x[2] > x[3] etc.....

But in reality the market changes, stretches/shrinks, amplitudes change, etc....


Here is the reality of the market, and we all look at the last 7 points and expect that our rule x[1] < x[2] & x[2] > x[3]... will work )).

As you can see, there is a pattern, it hasn't gone anywhere, but our cool 10-layer GPT-5 won't see anything ).


Conclusions: All AMO algorithms "out of the box" work with tabular data, i.e. all algorithms are looking for patterns with rigid binding to indices, it doesn't work....

1) It is necessary to look towards algorithms like associative rules, which are tied to events, sequence of events, but are not tied to indices.

2) Either design the right rules ourselves and make models from them

3) Or use for AMO such data representation which is used for algorithms of associative rules/sequences, but then the number of features will be billions.

 
Aleksey Nikolayev #:

It's not clear what kind of optimisation will work with them at all

What do you mean? There are continuous optimisation which is double and discrete optimisation which is integer. Or have I misunderstood?

GP is discrete, that's why there are problems with double
 

About the power of diversification

Let's say we have a CU that doesn't make very good money, not very good at all.

This is its yield curve.

In fact, it is a random noise with a very weak trend added, the trend is so small that it is not visible to the eye in the noise.

Here's the trend.

This is a strategy that we will not let trade.)

But what if we have 100 such uncorrelated strategies that are traded simultaneously on one account?

That's not very good, what if we have 1000 strategies?

And 100 thousand strategies?

That's pretty cool.

Is it possible to generate so many strategies with MO? ....

 
mytarmailS #:
What do you mean? There is continuous optimisation which is double and discrete optimisation which is integer. Or am I missing the point?

GP is discrete, that's why there are problems with double

I wrote there about algebraic data types. They generalise complex data types like lists and trees. They combine a complex discrete structure and a set of real numbers stored in this structure (it turns out to be of non-fixed size). Accordingly, we have to somehow combine discrete optimisation on the structure and continuous optimisation on the numbers stored in it. I have absolutely no idea how to do it at least theoretically.

 
Aleksey Nikolayev #:

I wrote there about algebraic data types. They generalise complex data types like lists and trees. They combine a complex discrete structure and a set of real numbers stored in this structure (it turns out to be of non-fixed size). Accordingly, we have to somehow combine discrete optimisation on the structure and continuous optimisation on the numbers stored in it. I have absolutely no idea how to do it at least theoretically.

I'm a bottom as a programmer, so I'm swimming in concepts.

What is the name of this combination? How to google it correctly?

1...264426452646264726482649265026512652265326542655265626572658...3591
New comment