Machine learning in trading: theory, models, practice and algo-trading - page 611

[Forester]

No one has shared an opinion on this:

Quote about the choice of the number of layers:

A network with three layers (numLayers=3: one input, one hidden, and one output) is usually sufficient in the vast majority of cases. According to Tsybenko's theorem, a network with one hidden layer is capable of approximating any continuous multidimensional function with any desired degree of accuracy. A network with two hidden layers is capable of approximating any discrete multivariate function.

I wonder if bar analysis refers to a continuous or discrete function?

I.e. for forex, maybe 1 hidden layer is enough or 2 should be enough?

From the last test, a simple model

88-14-2 proved to be better than the variants: 88-50-20-2, 88-32-7-2 (c this one is very close: 1-2%), 88-25-7-2, 88-32-2

[СанСаныч Фоменко]

elibrarius:

and the attributes and structure of the model turns out too

I keep reading and reading... and I can't understand: why the hell are you all hanging on to one of hundreds of model types - neural networks, in which, unlike the others, you need to understand very well the internal structure, which is not even simple?

Where and who has proven that the performance of a model depends in any significant way on its type?

What is the result of the NS? Does this result have any content? any interpretation?

[Forester]

Quote from here https://www.mql5.com/ru/code/9002 and accompanied by a picture:

Discrete is the third option - figures with gaps and voids.
If we transfer this to bars, then, for example, at small increments we must buy, at 25% of the maximum we must not, at 40% we must again, at 60% we must not and at 80% we must again buy.
It seems to me there is no such thing in Forex... and this is a continuous function. Although we have not only one trait, but many, and it's hard to imagine what kind of forces they create...
What do you think about it?

[Forester]

SanSanych Fomenko:

I keep reading and reading... and I can't understand: why the hell are you all hanging on to one of hundreds of model types - neural networks, in which, unlike the others, you need to understand very well the internal structure, which is not even simple?

Where and who has proven that the performance of a model depends in any significant way on its type?

What is the result of the NS? Does this result have any content? any interpretation?

I'll study one type thoroughly, when I'm done I'll look at something else. I don't like bouncing around on top of things.

[toxic]

Maxim Dmitrievsky:

http://www.valuesimplex.com/articles/JPM.pdf

[Maxim Dmitrievsky]

elibrarius:

No one has shared an opinion on this:

I wonder if bar analysis refers to a continuous or discrete function?

I.e. for forex, maybe the 1st hidden layer is enough or should I use 2?

From the last test, a simple model

88-14-2 turned out to be better than variants: 88-50-20-2, 88-32-7-2 (with this one very close: 1-2%), 88-25-7-2, 88-32-2

For example, the points are not distributed along a curve line but in groups separated by groups of other points... then this would already be a discrete f? Since it is impossible to connect all points with the 1st line

i.e. it is necessary to look at the scattering plots themselves along the line and think whether the 2nd layer is needed there or not

But so unreal to understand... I myself wonder how to understand what is better... I'll google later)

Ah, here's another thing... in regression, it always has to be continuous... and in classification, not necessarily... but I'm not sure

[Maxim Dmitrievsky]

I think so:

What a joke, read this document, Vasily recommended.

Thank you, I kept it, I will read it at my leisure... you can delete it :))) if necessary

or who cleans up your posts, you? :D

[Yuriy Asaulenko]

Maxim Dmitrievsky:

It's not like you're in the backcountry.

I do a lot of things that are unknown on this forum.)

Something may be known about, in due course.

[Maxim Dmitrievsky]

Yuriy Asaulenko:

I do a lot of things that are unknown on the forum ).

Something will be known in due course.

I switched to Wiener - I wonder... when these books will be over :) he basically tried to do forecasting as well.

[Yuriy Asaulenko]

Maxim Dmitrievsky:

I switched to Wiener - I wonder... when will these books be over :) he also tried to make predictions

He did not try, he was busy.) I am glad you liked it. You should read more.) In fact it saves a lot of time, you don't have to reinvent the wheel.