Machine learning in trading: theory, models, practice and algo-trading - page 3716
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
and what's in the input?
because "rubbish in, rubbish out", whatever the method.
Selecting, selecting, creating traits is another topic. Not less important, but not all at once.
But if someone wants to talk about traits, I personally don't mind and will support it as much as possible.
the topic about the MoD is good and useful, I guess :-))
---
why is everyone trying to predict specific price values ? No, of course it is understandable that it seems to be a shortcut to a golden island and a private jet.
But there are things that are potentially of significant practical use and can complement the classical TA.
For example, volatility - it is also predictable by usual methods plus or minus, but this plus or minus is missing.
If MO can predict tick volume with good accuracy (but based on historical prices, schedules, current time and news information), it is very, very cool.
To see a good forecast for the rest of the day 15 minutes before the Euro session starts is almost a poet's dream.
At least at least some result from the whole mash-education. By the way, there are no public confirmations of MO efficiency in trading.
To see a good forecast for the rest of the day 15 minutes before the euro session starts is almost a poet's dream.
At least at least some result from the whole mash-education.
Perhaps I should add a question - why does everyone on this forum try to forecast specific price values?
There are many models looking for volatility, starting from GARCH and further.
Well, when only prices (their increments) are searched for, the assumption of variance constancy is implicitly used, which is of course not always acceptable.
By the way, there are no public confirmations of MO effectiveness in trading.
What about the Medallion of the now deceased Simons? According to the common legend they rose on MO.
Oh, right, they forgot to put their algorithms in the public domain.
It is good that all funds without MO have put their algorithms in the public domain and we have the strongest evidence that NOT using MO guarantees efficiency.
probabilistic machine learning
The pros of probabilistic MOM (WMO) is that the model can provide more information. For example, as described above, you can get a prediction not only for the mean, but also for the variance. And for quantities with strongly non-Gaussian distribution (e.g. height of the knee of a zigzag), you can look at the shape of the distribution and its difference from what it would be at the SB.
A significant disadvantage of WMO is that the usual approach of learning with a teacher is lost. And in general it is much more complicated than standard approaches - you have to go into the underlying mathematics.
In the case of classification task, as I have already written, WMO approach does not give anything new, if algorithms giving probability of class membership are already used. But as we have already written in the thread earlier, the transition to the classification task is fraught with problems in terms of the meaning and trade logic of the algorithm.
Imho, the most developed area that can be attributed to WMO algorithms is survival analysis. I will try to further outline my opinion on how it can be applied to trading.
Survivability analysis studies the time of operation of some device until failure (life time of some organism). From a mathematical point of view it just means that a value with values about zero or more is studied. As for trading, such a value would be, for example, the maximum price movement towards a trade during the time from entry to exit. The height of the knee of a zigzag (minus the minimum step of the zigzag) is also suitable. But the price increment is not suitable, because it can be of different signs.
In survival analysis, instead of the value distribution function, the survival function, risk function or cumulative risk function are constructed. But all these functions are derived from the distribution function and you can always derive the distribution function back from them.
If (and when) I get my hands on it, I will give a simple example of using the approach.
If (and when) I get my hands on it, I'll give you a simple example of how to use the approach.
If (and when) I get my hands on it, I will give a simple example of how to use the approach.
I can't get to practice in the near future, so for now a small portion of theory.
What I see the point of WMO. I will refer to the approach from my article about gaps. There we were looking for distribution deviations of a certain value (specifically - maximum price advancement in the direction opposite to the gap closing). The next step is not just to look at the distribution of a value, but at its dependence on some predictors (signs, features).
Well, ok, we have built a model that gives the distribution of output by predictors, and then what? Further, if the obtained distribution is noticeably different from the one that would be at the SB, it indicates the possibility of trading. And then we use the same distribution to specify specific parameters of the entry/exit algorithm in this particular case. In the above-mentioned article, the distribution was used to determine the values of volume and stop loss, but they were the same for all inputs, and when using WMO, variability appears.
A little bit about what model I plan to use for the promised simple example. The first standard approach to using neural networks in trading is usually a model where a bar is predicted based on the previous N bars. By analogy, I plan to build the dependence of the distribution of the zigzag knee height on the previous N knees. In survival analysis, the Cox proportional hazard model is the best known, but in this case I think it is better to use parametric models.
I will say right away that I plan only to illustrate the method - there will be no requests from me to join the club of trillionaires or stakes on real)
I think I'll formalise the topic, and at the same time share a useful link https://link.springer.com/article/10.1007/s10959-022-01225-6. If you are interested, google the title and you will find it :-)
up