Machine learning in trading: theory, models, practice and algo-trading - page 587
You are missing trading opportunities:
- Free trading apps
- Over 8,000 signals for copying
- Economic news for exploring financial markets
Registration
Log in
You agree to website policy and terms of use
If you do not have an account, please register
No, it all came down to the fact that we should consistently build pseudostationary series, retrain as often as possible... in principle, that's what I do
or build linear/nonlinear filters... I understand that before this you need to consider the dynamics of change in the effect of predictors on the target, and try to adapt the output signal through the filter coefficients, depending on changes in the environment
Well, in general there is nothing special. At least in this chapter.
Once I discussed the problem of non-stationarity from the angle of machine learning models withVladimir Perervenko.
He convinced me that the problem of non-stationarity has nothing to do with MO. Since I never dealt with NS, I had no arguments to refute his opinion. Moreover, I had an intuitive understanding that various trees and others, except NS, work fine with non-stationary predictors.
Your post and the reference to your personal experience speaks on the opposite, at least in relation to NS it is necessary to consider non-stationarity of predictors.
If this is true, then there is almost the only tool available today that tries to work on non-stationary series - garch models. Filters, retraining on every bar... can not solve the problem of non-stationarity - it is guaranteed to fail, it will slip through the stop...
But the question remains about the other models, and there are a lot of them. I have no evidence that non-stationarity should be taken into account. The retraining in the models I have tried is always due to noise predictors.
Note that for me, solving non-stationarity and/or noise predictors are cornerstone problems in machine learning. The level of solving these problems determines the level of modeling error. The laboriousness of applying the models themselves is ridiculous and I do not take them into account.
If this is true, then today there is practically the only tool that tries to work on non-stationary series - these are garch models. ...
I would say that at the moment there is not a single tool that effectively predicts
There are individual cases - occasional temporary hits at a bull's-eye, when, in periods, you can earn good money
or the exploitation of arbitrage patterns, averaging
What I and you are doing - a system that would really predict in a meaningful way... It's a science fiction in terms of understanding all sorts of market processes and models :)
That's the wildest thing, people come here, read, and then panic and run to fill the grief of feeling helpless :)
I would say that at the moment there is not a single tool that effectively predicts
There are individual cases - occasional temporary hits at the bullseye, when, periodically, you can make a good profit
or the exploitation of arbitrage patterns, averaging
What I and you are doing - a system that would really predict in a meaningful way... It's a science fiction in understanding all sorts of processes and models :)
https://github.com/uber/pyro
Garch.... garch... Which garch are you talking about? - In R there are 24 of these garchs by keywords, by packages, probably about 12. Good and different).
Join
The rugarch package: ARMA(1,1); RealGARCH; Beveled t-distribution. Lots of intricacies
I would say that at the moment there is not a single tool that effectively predicts
There are individual cases - occasional temporary hits at the bullseye, when, periodically, you can make a good profit
or the exploitation of arbitrage patterns, averaging
What I and you are doing - a system that would really predict in a meaningful way... It's a science fiction in terms of understanding all sorts of market processes and models :)
I mean it's the wildest, people come here, read, and then panic and run to fill the grief from the feeling of their helplessness :)
I feel hopeless. I should take a break and go forward starting with Datamining.
I remember you had about 70% of adequate predictions. I wrote the post above.
I`ve tried to make my own predictions with MA (not simple, but gold, i.e. nonstandard). I should say everything was great at 70% of some time series, and I couldn't understand anything at the other 30%. The only thing is that there is no way to really use it.
Well, 70% of them are correct and that's nothing. Of those 70% correct to enter the transaction is about one-third. That leaves us with 23%. That's nothing against 30% of wrong predictions (and we don't know in advance if they are right or wrong). And the wrong predictions are in the inflection areas (the change in direction), and those areas are exactly the most suitable for trades.
On this basis, I believe it is futile to engage in prediction, and should be engaged in classification. I.e. to determine if a certain moment is suitable for a deal. Using models appears error of 20-40% entering. I gave more exact figures earlier in this topic.
Well, 70% of them are right, and that's nothing at all. Of those 70%, about one-third are correct to enter the trade.
Why a third?
All 70%. The forecast is valid for one hour. Then it works again.
It's futile to make predictions and you should classify them.
I don't understand anything.
For example, at 1 o'clock the predictor combination clause comes, which says that there will be a long in the next hour, i.e. until the next clause, until 2 o'clock in the afternoon.
How is it that you have a classification without a forecast? Why do you need any classification in financial markets, if it does not forecast?
How is it that classification exists without prediction? Why do you even need a classification in financial markets if it doesn't predict?