Machine learning in trading: theory, models, practice and algo-trading - page 2027

[[Deleted]]

mytarmailS:

Okay, sleepy,

I'm drunk on coffee, I can't sleep))

I'm not quite sure what to do with this bot... it depends on the seed in the tester. I don't really know what to do with it. it depends on the seed in the tester.

In real life it's like a random, i.e. I have to take the worst result and look for it... but there are drawdowns and so on. That's why I'm rooting for the rnn with the teacher, I want to do it, everything will be unambiguous there.

I want to play with this bot in real trade and now I'm wondering if it is worth it

[Aleksei Kuznetsov]

Maxim Dmitrievsky:

I'm not quite sure what to do with this bot... it depends on the seed in the tester. He sort of does not pour at any seed, but to get a good result, you have to pick up

In real life it's like a random, i.e. you have to take the worst result and look for it... but there are drawdowns and so on. That's why I'm rooting for the rnn with the teacher, I want to do it, everything will be unambiguous there.

If I want to try it out in real trading, I may start a real trading robot, but I don't think it is worth it.

Maybe 5-10 models with different Seed to average? It will be an average between the best and the worst.

[[Deleted]]

elibrarius:
Can we average 5-10 models with different Seed? It will be an average between the best and the worst.

In the tester, the timer works one way, in real life another.

The network is retrained all the time, there is an element of chance. Before each retraining plug in

[Aleksei Kuznetsov]

Maxim Dmitrievsky:

And in the real world, what's the sideline? In the tester, the timer works one way, but in the real world it works another way.

The network is constantly retrained, there is an element of chance. Before each retraining, plug in

Averaging with 5-10 models will reduce dependence on the sid, both in the tester and in the real world, and the method of its calculation is not important (obviously from the current time), averaging is important.

In general it is better if the Sid is either absent in the model or has almost no effect on the result.
If it does, we get additional element of noise and instability to other very noisy input data.

[Aleksei Kuznetsov]

Maxim Dmitrievsky:

it'stoo slow and there will be fewer signals

left it as it is

like this for 3 months, then flies away immediately in trade, after testing. I.e. the same model kind of continues to work

If I wanted to use the real time version I would have to re-calculate the real time version.

If you want to be sure of the correct order, check it once again and look at the real order.

[[Deleted]]

elibrarius:

Additional checking is never superfluous, especially at the risk of losing real money.

Model averaging leads to a significant reduction in the number of signals, because many will be contradictory.

Unless you run them in parallel

[Andrey Dibrov]

Ah, do you have practice?

[Valeriy Yastremskiy]

Kharitonov's Economophysics. I read it, I like it).

[Rorschach]

Valeriy Yastremskiy:
Kharitonov's Economophysics. I read it, I like it).

I likedthis one .

[Aleksey Nikolayev]

Valeriy Yastremskiy:
Kharitonov's Economophysics. (I read it, I like it).

The basic area of economophysics (potential game theory) does not seem to be reflected there at all.