Machine learning in trading: theory, models, practice and algo-trading - page 3442

[Rorschach]

Maxim Dmitrievsky #:

What if we remove increments from the sample that do not cover the specified spread.

Or some other filtering methods.

That doesn't make sense to me. I sent you my idea the other day

[Maxim Dmitrievsky]

It's amazing what similar models we were able to train, with very significant randomisation within the algorithm

The kozul just got even more interesting.

Learn 0 model
R2: 0.9609434010382693
Learn 1 model
R2: 0.9619192065934826
Learn 2 model
R2: 0.9533013034947937
Learn 3 model
R2: 0.96256861996626
Learn 4 model
R2: 0.9087029748259404

[Maxim Dmitrievsky]

Rorschach #:

It doesn't make sense to me. I sent you my idea the other day

Didn't remember it, I guess.

[mytarmailS]

Maxim Dmitrievsky #:

It is surprising what similar models we were able to train, with very significant randomisation within the algorithm

The kozul just got even more interesting.

It is standard to train on a training sample, tune on a validation sample and test on a test sample of one model.

Do I understand correctly what you are doing in your kozul is :

you train several models on different parts of the test sample and they must all pass the test sample.

Is that all, or are there more details?

[Maxim Dmitrievsky]

mytarmailS #:

The standard is to train on a traine sample, tune on a validation sample , and test on a test sample of one model.

Do I understand correctly what you are doing in your kozul is :

you train several models on different parts of the test sample and they must all pass the test sample.

Is that all, or are there more details?

You've described the CV. Only you don't train on the test sample by definition. It's a test sample 🫥

[mytarmailS]

Maxim Dmitrievsky #:
You CV described it. Only a test sample isn't used for training by definition. It's a test sample 🫥

Oops. I meant train. I was sleepy.

So what's the difference with CV?

[Maxim Dmitrievsky]

mytarmailS #:
Oops. I meant Train. I was sleepy.

So what's the difference with the CV?

Well, it's used to get unbiased estimates. Like, you add a new variable, you have to look at its effect. With one model there will be bias in the estimation, with cv there won't be.
.

By analogy, I look at how model training affects the TC on average. That is, in what places the model is wrong on average, and in what places it works. To separate where it can be traded and where not. It can be used in different ways.

And then I am already trying to separate these classes into trading and non-trading, with low error. And everything is not perfect there yet. The brains are not enough.

[Maxim Dmitrievsky]

mytarmailS #:

I remember we once had a useful argument about HMM and K-means, that they sort of cluster the same or not the same....

do you remember what it was about? To identify market modes.

[mytarmailS]

Maxim Dmitrievsky #:

I remember we once had a useful argument about HMM and K-means, that they sort of cluster the same or not the same.

do you remember what it was about? To identify market modes.

Probably just prices with a sk. window but I can't say, I don't remember.

[Maxim Dmitrievsky]

mytarmailS #:

Probably just prices from the window, but I can't say for sure, I don't remember.

Nowadays it is common to add the word "causal" to everything - and it reads beautifully and with a hint of magic :)

https://github.com/LilJing/causal_hmm?tab=readme-ov-file