Machine learning in trading: theory, models, practice and algo-trading - page 2730

New comment
 
Aleksey Vyazmikin #:

You can only mix within a sample, if you mix two samples, you are denying that the market is changing.

Can't you see the logic again?
It is pointless to compare series to determine the optimal length of the training sample, because the market is changing

You can mix them at any point, it won't change anything.
 
Maxim Dmitrievsky #:
Can't you see the logic again?
It is pointless to compare series to determine the optimal length of the training sample, because the market is changing

You can mix at any point, it will not change anything

How can you prove that the market is changing? How long does this process take? Or is it constantly changing?

 
Aleksey Vyazmikin #:

How can you prove the changeability of the market? How long does this process take? Or is it constantly changing?

That's it, the delightful arguer is on.
Non-stationarity at the very least. Constantly changing, sometimes stopping for a smoke break.
 
Maxim Dmitrievsky #:
That's it, the delightful arguer is on.
Unsteady at the very least. Constantly changing, sometimes stopping for a smoke break.

And what sample size should be taken to determine stationarity/non-stationarity?

According to you a pattern lives no longer than the lifetime of the sample change, but what if I have a pattern in my sample that repeats for 8 years? What is that, an anomaly, or the patterns are not all changing or the patterns identified in a small area are erroneous and due to other factors?

 
Aleksey Vyazmikin #:

Each predictor individually is a numerical sample, so why not estimate them individually and average the result?

This works only in the case of independent features, and since they are counted at the same price, it is not possible. In the case of dependence everything is much more complicated - we can take copulas as an example, where univariate distributions are always the same uniform, but bivariate distributions can be very different.

Aleksey Vyazmikin #:

Perhaps we should find those variants that will give the best results in terms of identifying the belonging of segments to a particular group and the efficiency of training on a grouped population.

You have an appetite for heavy enumeration calculations) We will have to add (to the already considerable amount of enumeration) enumeration by feature types and, probably, by feature parameters.

Nevertheless, it seems to me that there is a rational grain in your approach, there is something to think about.

 
Aleksey Vyazmikin #:

Didn't I write that the idea is to compare samples (training and application), that if your theory is correct, the sample will cease to be similar as it increases, and in order to understand this we need criteria for assessing its change, which are derived from the methods of assessing similarity?

Maybe instead of statistical criteria of sample homogeneity you should just watch the change of feature importance of the model in dynamics (in a sliding window).

If there is a strong discrepancy between the current state and the previous state, it means that we are already in a different sample.....

Pros:
1.You don't need to programme stat. Tests, everything is ready out of the box.
2. It takes into account not only the change in time of the sample, but also the change in the target sample, which I think is not less important.

 
Aleksey Vyazmikin #:

And what sample size should be taken to determine stationarity/non-stationarity?

According to you, a pattern does not live longer than the lifetime of a sample change, but what if I have a pattern in my sample that repeats for 8 years? What is that, an anomaly, or the patterns are not all changing or the patterns identified in a small area are wrong and due to other factors?

I would say no more than the lifetime of a particular trend on an arbitrary time scale
But that's a loose description.

From point to bifurcation point.
 

Different models but similar, different and not similar how do they differ? The bifurcation point will not necessarily lead to a change of the model, it is possible to mark the same areas visually manually, but there is no predictive part at the end, the goal is to find the minimum length of the sample, which confirms the state or compliance of the model.

Complexity of the model, here of course there is also a contradiction, a simple model will not describe a sufficiently necessary long section, but will be repeated, a complex model can describe a sufficiently necessary section in length, but may be unique. As always something in the middle is needed))))))

 
Valeriy Yastremskiy find the minimum sample length that confirms the state or model fit.

Complexity of the model, here of course there is also a contradiction, a simple model will not describe a sufficiently necessary long section, but will be repeated, a complex model can describe a sufficiently necessary section in length, but may be unique. As always something in the middle is needed))))))

In general, different models differ in their implementations of randomness ) and are similar for the same reason

Especially when tens and hundreds of features are used. Some of them work on the forward, some do not. But there is no way to select them.

Only with the help of a good moonshine plant, as suggested above.
 

It's an unimaginable mess: everything is mixed up - horses, people.....


We can distinguish two types of models

1. Based on machine learning ideas.

2. Statistical models, which are fundamentally more widely used in financial markets.


MO

As it seems to me all MO algorithms have one goal - to find some number of patterns. In this case, a pattern is a string with a teacher value and feature values. There is no value of rows next to each other! The number of such patterns can be looked up in RF, from about very often with 50 trees the fitting error changes very little. More than 150 trees is meaningless. That's the diversity of financial markets.

And we should reason about the lifetime of these trees, which (lifetime) is determined by the stability of the connection between the features and the teacher. I.e. we should deal with the connection between traits and teacher.

1...272327242725272627272728272927302731273227332734273527362737...3656
New comment