Machine learning in trading: theory, models, practice and algo-trading - page 2561
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
xgboost can also do it, but it's very hard to write your own function. You have to derive formulas.
http://biostat-r.blogspot.com/2016/08/xgboost.html - 6th paragraph.
Yes, you'll have to calculate on paper) Also not every function will do - you need to define the second derivative, which probably must be non-zero.
Right, that's exactly the relationship between the predictor and the target I'm talking about.
Now, I am not aware of a method for building a model that gives an estimate of "stationarity" at different sampling intervals with splitting or some other mechanism for combining predictors. All models do a fit to sample plots, estimating only a quantitative measure of improvement, but you need to estimate it across intervals, then the model might be more robust.
Why not just do a cross validation to check stability by intervals?
Here's that article, but it's not regularity, it's irregularity. But whatever.
https://www.ncbi.nlm.nih.gov/pmc/articles/PMC518821/
I think not - an example in python.
No, it's not that, just a custom metric...
you can't put ffs in there.
You see, with a custom metric you give the target and the data and choose the formula by which to read the error.
And with FF you can say - AMO! I don't know what should be the targeting and how to do it better, but make it good (and you can do it by many criteria at once)
And ff + AMO will "come up with a target," do the fitting, etc..
This is a different world...
why not just do a cross-validation to test for interval stability?
It's not about finding a randomly successful model, but about increasing the probability of making this model successful.
No, it's not that, it's just a custom metric...
You can't put a ff in there.
You see, with a custom metric you give the target and the data and choose the formula by which to read the error.
And with FF you can say - AMO! I don't know what should be the targeting and how to do it better, but make it good (and you can do it by many criteria at once)
And ff + AMO will "come up with a target" themselves, adjustments to do, etc..
It's a different world...
Perhaps this is something in the spirit of Ivakhnenko, in any case - not my thing.
I read a book from the 70's, and it says that if there is no autocorrelation, then no prediction is possible. Is there anything more modern on this subject?
I guess it depends on the choice of model, for the type of process to be assumed.
After all, in static processes, on the contrary, they struggle with autocorrelation with linear models.
While in dynamic processes there is mainly autocorrelation,
this is why the scientific community tries to solve the problem of autocorrelated processes with appropriate models.
Hence probably the statement that if there is no autocorrelation, then the forecast with large errors.
That is, one must distinguish the nature of the process, and estimate it with suitable algorithms.
I read a dissertation from 2014 on synthetic estimation of dynamical systems.
The paper looked at examples with an exponentially correlated process.
Is it a coincidence? I don't think so.
Perhaps it is something in the spirit of Ivakhnenko, in any case - not mine.
No, Ivakhnenko has nothing to do with it, there are just some problems that cannot be solved with a target, there is no target ready-made.
Imagine the problem:
There are attributes and there is a price, all of which is a matrix of attributes "X",
problem:
I want AMO to take "X" as input and give a function on the output that :
1) repeat the price as much as possible (correlate)
2) Will be in the range -1,1
3) outperform the price as much as possible (negative cross correlation)
Everything. Let's take any optimizer, in this case a multicriteria one (Pareto) and start playing with the (guts) of a honeycomb until it gives us what we need in the output
You see, it's impossible to have a ready-made targeting, there's only search and run through a fitness ...
How do you optimize your network for max profit or recovery factor, or both? How do you do it as a ready-made target?
No, Ivakhnenko has nothing to do with it, there are just some problems that cannot be solved with a target, there is no target in a ready-made form.
Here's a problem for you:
There are signs and there is a price, all of which are a matrix of signs "X",
problem:
I want AMO to take "X" as input and give a function on the output that :
1) repeat the price as much as possible (correlate)
2) Will be in the range -1,1
3) outperform the price as much as possible (negative cross correlation)
Everything. Let's take any optimizer, in this case a multicriteria one (Pareto) and start playing with the (guts) of a honeycomb until it gives us what we need in the output
You see, it's impossible to have a ready-made targeting, there's only search and run through a fitness ...
Or just a simple example to optimize the network to trade on the maximum profit or recovery factor, or both at once, how to do it as a ready-made targeting??? there is no way!
Well, multi-criteria optimization. You can either combine one compromise out of all of them, and then watch how the solution changes when you change the weights of the criteria and choose the best of them on the forward. Or you can select one criterion as the main, and for others select admissible limits and add a hard penalty for breaking them and also watch which variant is better.
If you get a serious advantage on the forward test, then it makes sense to try, otherwise it'll be just complexity for the sake of complexity.
Offtop, but fun
https://www.youtube.com/watch?v=_Aow6P3oBAg