Machine learning in trading: theory, models, practice and algo-trading - page 618

 
Aleksey Terentev:
Then input 100 bars permanently. Model of neural network will be as follows: input - 100, hidden - x, output - 5.

in this case it will turn out that the model from the static 100 bars will be neutered, and I believe that this will not lead to the desired search for a possible pattern (

 
Anatoly Zainchkovskii:

Maxim, you're running your neural networks on monopair, aren't you? Have you ever thought about the fact that you can create a convenient row, which may then show itself better on the forward? After all, in fact, let's say a headshoulder figure, for example, does not occur very often, but imagine that you can do it every hour...


If you've got a good idea, it may be a good reason to buy an index, but I don't think that it has any effect on the market.

 
Maxim Dmitrievsky:

Yes, I was just building a portfolio, first with regression, then I wanted to use NS, but I didn't finish... it gets the same non-stationarity, it's hard to pick up the instruments... Mostly you need to trade indices with this strategy

Maxim, what are you feeding to the input of the neural network? Koldun gives increments as input, what about you?
 
Maxim Dmitrievsky:

Yes, I was just building a portfolio, first through regression, then I wanted to use NS, but I didn't finish... it gets the same non-stationarity, it's hard to choose the instruments... mostly you need to trade indices using this strategy


I may be thinking that it's because I've added a simple cycle to increase the length of the model. Now I get a good image at any time. But forward waters have the same 50/50 and I'm trying to re-sort them...

 

By the way, besides the increments themselves I would also feed the time of each bar as a qualitative parameter...

 
Anatolii Zainchkovskii:

By the way, in addition to the increments themselves, I would also feed the time of each bar as a qualitative parameter...

That's exactly how it should be done. The sample volume should be calculated from probability theory.
 
Alexander_K2:
That's exactly how you should do it. And the sample volume should be calculated from probability theory.

I do not understand about the volume, for training is not enough 10,000 examples of states?

 
Anatolii Zainchkovskii:

I do not understand about the volume, for training is not enough 10,000 examples of states?

Yes, it is. But you have to calculate for each pair separately. They are different dogs. Probability density functions and amplitudes are very different.
 
Alexander_K2:
Quite. But you have to calculate for each pair separately. They are different dogs. Probability density functions and amplitudes are very different.
And why count - take 20,000 in reserve, for example, and that's it!
 
Alexander_K2:
Quite. But you have to calculate for each pair separately. They are different dogs. Probability density functions and amplitudes are very different.

I think in my portfolio approach there is no need to count pairs separately... just take the increment of the portfolio itself and time points...