Machine learning in trading: theory, models, practice and algo-trading - page 595

[Grigoriy Chaunin]

SanSanych Fomenko:

When you enter a building called "Statistics," it says"Garbage in, garbage out" above the entrance.

Any indicator is a function of price. Neural networks have a nonlinearity. It is capable of deriving any indicator formula by itself if the network is deep enough. If the network cannot learn on price data, it is not because of inputs but because of the fact that it is impossible to obtain output data on price.

[Yuriy Asaulenko]

Grigoriy Chaunin:

Any indicator is a function of price. Neural networks have a nonlinearity. It is capable of deriving any indicator formula by itself, if the network is deep enough. If the network cannot learn from price data, then it is not about the inputs, but about the fact that you cannot get output data from price in principle.

Both you and SanSanych are right.

On the one hand, the NS will automatically build the necessary indicator and their combination. On the other hand, if the data is not clean, and there is too much noise in it, no NS will be able to learn anything. So, it's about the inputs too.

[Maxim Dmitrievsky]

How important is it to mix the sample when training the NS? What are the mathematical justifications for this?

Is mixing relevant for all MO models or only for some specific ones?

[Yuriy Asaulenko]

Maxim Dmitrievsky:

How important is it to mix the sample when training the NS? What are the mathematical justifications for this?

Is mixing relevant for all MoM models or only for some specific ones?

Mixing is necessary so that the learning algorithm doesn't follow the same path every cycle. We can get there and fail to get out of the local extremum. But mixing is not a panacea either.

[Maxim Dmitrievsky]

Yuriy Asaulenko:
It is necessary to mix it up so that the learning algorithm does not follow the same path with every cycle. We may get into and not get out of the local extremum.

I.e., you need to mix a few times, teach a few times and see how the results correlate?

[Yuriy Asaulenko]

Maxim Dmitrievsky:

I.e., do you need to mix and train a few times and see how the results correlate?

You need to shuffle after every few epochs of training. Unfortunately, many learning algorithms do not allow breaks (see Python - some packages (modules)) and start from scratch every time.

Stirring is also good to combine with annealing. But, again, it is difficult to do it on automatic. You always need to watch the intermediate results, and then plan your next steps.

[Maxim Dmitrievsky]

Yuriy Asaulenko:

It is necessary to shuffle after every few epochs of training. Unfortunately, many learning algorithms do not allow breaks (see Python - some packages (modules)) and start from scratch every time.

Stirring is also good to combine with annealing. But, again, it is difficult to do it on automatic. Because you always have to watch the intermediate results, and then plan further actions.

wow... that's it... i.e. just mix it up before training makes no sense

[Vizard_]

Alexander_K2:I know people who pulled the grail from NS, but those guys are so closed to communication and even hints about what they do, I'm a beginner so definitely have no chance. I only know that everything is complicated, it's not Vels, not Metatrader and not even S#, but C++ and MatLab with some chips that decode and interpret the data coming from the caliders, it turned out that this is the same methodology, I heard it and got scared, they work with uncle who used to grind terabytes a day at CERN, looking for new particles in quantum chaos.

And you've got it now) The rattle is CatBoost.
---------
If you ever feel like fishing for a boson...
https://www.kaggle.com/c/higgs-boson

[Forester]

In Darch, the default is to mix before each epoch. I tried to turn it off - it didn't learn anything at all.

So I thought, if everything is shuffled, how can I make the fresh data have a stronger effect on learning?

[Aleksey Terentev]

elibrarius:

In Darch, the default is to mix before each epoch. I tried to turn it off - it didn't learn anything at all.

So I thought, if everything is shuffled, how can I make the fresh data have a stronger effect on learning?

Try cross-validation (K-fold).