Machine learning in trading: theory, models, practice and algo-trading - page 2235

[Aleksey Nikolayev]

Valeriy Yastremskiy:

Schrödinger has long failed, today I read his review on the dissertation about numerical simulation of nonlinear spin waves in graphene structures on matlab, of course Schrödinger is used, but already noting that he does not cope)))

This is probably a hobby at home) In general for a long time surprised, the academy of finance like, and talk about quantum computers, CERN, Dubna) But today surprised just) The child (my) in graduate school, he asked a question, and the network in Python writes?) That's countered and asked)

Earlier I heard that in mathematics in Russia now the Physics and Mathematics Department (Higher School of Economics) is in the lead, while the Faculty of Mechanics of Moscow State University has lately given up.

[Valeriy Yastremskiy]

Aleksey Nikolayev:

I've heard in passing that in mathematics in Russia now the Physics Department and the Higher School of Economics are at the top of the list, while the Faculty of Mechanics at Moscow State University has fallen drastically lately.

Off the subject, but in general, it's bad). And among the sheep .... It seems that Phystech never gave up its position. I am not very familiar with HSE and MSU, a couple of acquaintances is not an indicator)))

[iwelimorn]

Maxim Dmitrievsky:

Seriously... this weirdness is breaking my brain. Without the spread, it works well on the track/test.

With the spread, it works well on the track, but it breaks on the test. What's so different on the test that the spread tick by tick prevents it from turning into a profit...

Seems to be some kind of bug in the logic.

Last night I added one more "do not trade" class to the target:

    for i in range(dataset.shape[0] - max):
        rand = random.randint(min, max)
        curr_pr = dataset['close'][i]
        future_pr = dataset['close'][i + rand]

        if future_pr  - curr_pr < -25*POINT:
            labels.append(1.0)
        elif future_pr - curr_pr > 25*POINT:
            labels.append(0.0)
        else:
            labels.append(2.0)

Rewrote tester, in the end it does not give anything even a little decent result. And the culmination for me was that the save_model method for C++ doesn't support multiclass models, dog. In short, based on the conditions it's a dead end.

[Maxim Dmitrievsky]

welimorn:

Last night, added another "do not trade" class to the target:

rewrote the tester, in the end nothing gives even a little decent result. And the culmination for me was that the save_model method for C++ doesn't support multiclass models, dog. In short, based on the conditions, a dead-end branch .

Over another classifier that allows/prohibits trading

but that really doesn't make much sense, I checked it out )

[Aleksey Vyazmikin]

There are also such optimization solutions.

[Aliaksandr Hryshyn]

Aleksey Vyazmikin:

There are also such optimization solutions.

I should take a look, maybe there will be something useful

[mytarmailS]

Aleksey Vyazmikin:

There are also such optimization solutions.

Why are you interested in optimization?

[Aleksey Vyazmikin]

mytarmailS:

Why are you so interested in optimization?

This is instead of gradient error correction. It's true that a larger number of features requires more computing power, but maybe for collapsed features it will do...

[mytarmailS]

Aleksey Vyazmikin:

This is instead of gradient error correction. It is true that a larger number of features requires more computing power, but it may work for collapsed features...

Are you training a network?

[Aleksey Vyazmikin]

mytarmailS:

Are you training a network?

Boosts also use a gradient. This is just information to expand knowledge and methods suitable for MO.