Machine learning in trading: theory, models, practice and algo-trading - page 3636

New comment
 
Andrey Dik #:
Take the first plot as in the picture first

I have already written above that on non-representative data there will always be bad models (more precisely, there will be a small chance to get a satisfying full process). We deal with such data in the market.

 

Forum on trading, automated trading systems and testing trading strategies

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

fxsaber, 2024.11.17 12:27 pm.

100% there have been studies on which analytical functions are poorly amenable to MO.

I guess there is no point in reinventing the wheel. And search to find a lot of information on the topic.

[Deleted]  
There is no point in predicting some piece of some function because different phs may need their own preprocessing, this is not an MO problem in pure form but a preprocessing problem.
 
fxsaber #:

Unfortunately, it is not obvious to me from the analytical form of the function that it is periodic. But if several periods fall into the training interval, even a human can predict its behaviour. That is, it is not interesting to take such a learning interval at all.

It is much more indicative to take an interval, for example, two times smaller than the period, but without restrictions on the number of training points.


That's what I'm talking about, you've got the point right, I mentioned it above).
 
Aleksey Vyazmikin #:

I have already written above that on non-representative data there will always be bad models (more precisely, there will be a small chance to get a model satisfying the full process). We deal with such data in the market.

This is what we were talking about above. The problem on real problems is always that there is an unknown part of the process where it is reliably known that full periods are contained. I gave such a simple example, where the full period of a function is unknown, but prediction of the next segment after the training one is required. And, I said above that even in this simple example, there is such a set of network that will be able to predict the next segment of the process. The point is to find that right set of network. MO methods do not solve this problem reliably, hence the problems on new unknown data.

This is why amateur MO amateurs like to use the term "non-stationarity" with respect to DEM in case of failures, but even a simple example with a completely known formula will get problems.

[Deleted]  

It seems that optimisers start optimising f-i in case of failure on price data, and the term non-stationarity was known before their attempts )

it's all because of the mess in their heads and incorrect application of tools.

An excellent example is methods of processing cypheric signals that are inapplicable for non-stationary processes.

Another excellent example is optimisation of left-hand functions that are not relevant to trading

 
Maxim Dmitrievsky #:
The function is periodic and stationary. Easy to predict.

Maxim Dmitrievsky #:
There is no sense in predicting some piece of some function, because different phs may require their own preprocessing, this is not a pure MO problem, but a preprocessing problem.

First you "easily" solved the problem, and then claimed that there is no point.))

What prevented that dancer from using preprocessing in the problem is known only to the dancer.😁

[Deleted]  
Andrey Dik #:

First "easily" solved the problem and then claimed there was no point.))

Who prevented that dancer from using preprocessing in the problem is known only to the dancer.😁

I added representative data and everything is predicted. But there is no point in predicting on a limited data set. Simple preprocessing was used. Questions?
So the understanding of how MO works doesn't come through?
[Deleted]  
Does the model predict? It does. Questions?
 
Andrey Dik #:

This is what we were talking about above. The problem on real problems is always that there is an unknown process segment where it is known for sure that the full periods are contained. I gave such a simple example, where the full period of a function is unknown, but prediction of the next segment after the training one is required. And, I said above that even in this simple example, there is such a set of network that will be able to predict the next segment of the process. The point is to find that right set of network. MO methods do not solve this problem reliably, hence the problems on new unknown data.

I think you realised that the members of this thread understand this very well?

So what method do you propose to solve your problem reliably?

1...362936303631363236333634363536363637363836393640364136423643...3741
New comment