Machine learning in trading: theory, models, practice and algo-trading - page 3353

New comment
 

Don't fight, guys, we're reading you.

open the locks ;)


 
СанСаныч Фоменко #:

There is no other way of thinking! We use ready-made MO algorithms that are accompanied by a set of additional functions. Everything together is called a "package".

What are"real class probabilities"? For example, the function

returns"probability class estimates". No other probabilities other than "estimates" the algorithm can contain.
The question is not about what it can. It's about how to get reliable class probabilities. So that you could be sure that with a class probability of 0.8, 80% of cases were predicted correctly. And you could use a threshold, for example. The classifier output does not do that in most cases, I repeat again. They either overestimate or underestimate "by design". That's why the threshold doesn't work. Real probabilities are when they neither overestimate nor underestimate.

You've already shown that you don't know. So there is more to learn. So "we need to master all MOEs" and get rid of batch thinking.
 

It seems that it is not about point estimation of probability, but about its interval estimation. For matstat, this is a common approach - not just to obtain a specific numerical estimate of probability, but also to obtain an interval into which the true value of this estimated probability falls with a given accuracy (probability). Here there is some difficulty in understanding, because the concept of probability participates in two different hypostases - both the estimated value itself and the accuracy of its estimation. And these are quite different probabilities)

Although I have not studied conformal forecasting in detail, I may be wrong.

 
Maxim Dmitrievsky #:
The question is not about what he can do. It's about how to get reliable class probabilities. So that you can be sure that with a class probability of 0.8, 80% of cases are predicted correctly. And you could use a threshold, for example. The output of the classifier is not true in most cases, I repeat again. They either overestimate or underestimate "by design". That's why the threshold doesn't work. Real probabilities are when they neither overestimate nor underestimate.

That's not what you have. The 0.8 figure quoted is one of the class probabilities. Here's a histogram of the class probabilities.


And I have it exactly like that and no other way, because if it is otherwise, it means overtraining. For me, at a fixed threshold, the mismatch of prediction error on the OOV and OOS and on the VNE file is the main sign of overtraining. I have the threshold working fine. And "real probabilities" is from the realm of some fiction that has nothing to do with the real code and terminology used in this case.

 
СанСаныч Фоменко #:

You have it wrong. the 0.8 figure given is one of the class probability values. here is a histogram of the class probabilities.


I have it exactly like this and no other way, because if it is different, it means overtraining. For me, at a fixed threshold, the mismatch of prediction error on the OOV and OOS and on the VNE file is the main sign of overtraining. I have the threshold working fine. And "real probabilities" is from the realm of some fiction that has nothing to do with real-world code and the terminology used for it.

How did you realise that your threshold works perfectly?
For you it's fantasy, and for someone else it's commonplace.
 
Aleksey Nikolayev #:

It seems that it is not about point estimation of probability, but about its interval estimation. For matstat, this is a common approach - not just to obtain a specific numerical estimate of probability, but also to obtain an interval into which the true value of this estimated probability falls with a given accuracy (probability). Here there is some difficulty in understanding, because the concept of probability participates in two different hypostases - both the estimated value itself and the accuracy of its estimation. And these are quite different probabilities)

Although I have not delved into conformal forecasting in detail and I may be wrong.

We are talking about a slightly different approach, before anyone googled it :)
 
Maxim Dmitrievsky #:
How did you realise your threshold was working perfectly?
For you it's fantastic and for some it's commonplace.
Matching prediction error on the OOV and OOS and on the INE file
 
СанСаныч Фоменко #:
Matching prediction error on the ALE and OOS and on the SNE file
How did you realise that the classifier gives the correct probabilities? Not just the values in the range. Are you reading what is being written to you?

If you set a threshold of 0.8, will 80% of the trades be profitable? And if it's 0.51?

It's almost certain that you won't have that. Check it out.
 
Maxim Dmitrievsky #:
How did you realise that the classifier gives the correct probabilities? Not just the values in the range. Are you reading what is being written to you?

Probabilities of models are given by statistics on the training sample.

Accordingly, without a representative sample they are not accurate, so get over it :)

Either figure out what the model consists of, and reweight the leaves according to the algorithm you've devised...

 
Aleksey Vyazmikin #:

The model probabilities are given by the statistics on the training sample.

Accordingly, without a representative sample they are not accurate, so get over it :)

Either figure out what the model consists of and reweight the leaves according to the algorithm you have devised....

The probabilities of the model are given by the sigmoid, not this. For simplicity's sake, you take the track and the shaft, no matter what's outside. And even there, you get a stutter.
1...334633473348334933503351335233533354335533563357335833593360...3536
New comment
Reason: