Machine learning in trading: theory, models, practice and algo-trading - page 2174

 
Alexander_K:

Yeah...

The subject gets better every day.

And, after all, it seems to be simple. There are coryphaei of MO here, but practically, there are no wizards who can offer some magical features (right?), which would have predictive ability. Proving theoretically that these are really great chips, of course...

Now, it has been said, not by me, but by everyone, that the simplest model for prediction is the sign of the next return of a stationary process with a non-zero ACF.

The ACF is lucky - in the market, no matter what you do with the quotes, it is always nonzero and this gives a wild hope in the hearts of those who suffer.

But where is stationarity and how can we predict the sign of the next increment if it may be "stupid" = 0?

Ahem... I've already shown a passage from the Book of Genesis many times. I'll show it again:

"... everything is elementary, the stationary distribution is obtained by writing bars with equal number of ticks.
at that the distribution of time intervals between OPEN of such bars is exponential:

, and the distribution of increments is double triangular:

Now, having such a distribution of returns, predicting the sign of the next one does not seem to be such an impossible task.

Ahem...

Up to 4 ticks will do? OHLC?
 
Renat Akhtyamov:

It's just that the conditions are steep.

the commission is 6 times the spread

Not any system will fit there

;)))

Ok, so it's eva

I'll run it on at least cents so as not to crap on the market, let's say ;)))

but you can do it on the demo, i don't care

it means such a result because of the commissar, the spread is small

 

Thank you very much! I tried to use the principles of autoencoding with simple full-knowledge neural networks, but not quite successfully.) Maybe I should reconsider the experience)))

I don't know where the code is, because it's referenced and I don't know where to look for the implementation.

Of these 20, we limit the analysis to 16 for which we have or can approximate the relevant inputs. The notebook conditional_autoencoder_for_trading_datademonstrates how to calculate the relevant metrics.
 
Dear, esteemed MOshnikov, have you tried to predict for two 3, 5 10 steps ahead. Or only on history? Well I do not see any post with the results, even from the demo. Sorry I forgot. "Spitfire" has shown negative results. Twice. Hee hee.
 
Aleksey Vyazmikin:

And where is the code there, because it is referenced, but it is not clear where to look for the implementation.

Of these 20, we limit the analysis to 16 for which we have or can approximate the relevant inputs. The notebook conditional_autoencoder_for_trading_datademonstrates how to calculate the relevant metrics.

There is something in there though.

https://github.com/0b01/recurrent-autoencoder

0b01/recurrent-autoencoder
0b01/recurrent-autoencoder
  • 0b01
  • github.com
Dismiss Join GitHub today GitHub is home to over 50 million developers working together to host and review code, manage projects, and build software together. Sign up GitHub is where the world builds software...
 
Uladzimir Izerski:

Sounds like you've had a few beers, brother. And beer is bad for your thinking centers.

1. You're not my brother.

2. I didn't drink beer, you're a booze drinker. Go get some cognac for your think tanks. Hee hee

 

Still, there's something there.

https://github.com/0b01/recurrent-autoencoder

there is a good package on variation, I want to use it. It's on top of PyToch.

https://pyro.ai/examples/index.html#

Getting Started With Pyro: Tutorials, How-to Guides and Examples — Pyro Tutorials 1.5.0 documentation
  • pyro.ai
Welcome! This page collects tutorials written by the Pyro community. If you’re having trouble finding or understanding anything here, please don’t hesitate to ask a question on our forum! New users: getting from zero to one¶ If you’re new to probabilistic programming or variational inference, you might want to start by reading the series . If...
 
elibrarius:
To 4 ticks will do? OHLC?

I already answered - a certain Demco has formed equal-tick bars (100 ticks per bar according to Alpari) and has worked with OPEN prices of such bars.

In fact, the initial flow is thinned and the simplest flow with such an unusual distribution of increments is obtained. Demko called it a "progenitor" of market processes.

I personally checked his data - indeed, a stationary process is obtained in such increments.

The question is: why don't I take these increments, study neural networks and make a secret accession to the Grail?

The answer is, I don't know. I have some kind of working TS, and I'll leave this option for later sometime...

 
Maxim Dmitrievsky:

All the same, but passes the test from 2017.

Raised lot, 3.5k% in 3 years. Max balance drawdown 10%, absolute 23%. Z.I. You can do even better. Python sources will be in the article.

Enjoy

Bomb!
 
welimorn:
Bomb!

The proper addition of the spread to the signs should improve it even more. I'll look into it this week.