Machine learning in trading: theory, models, practice and algo-trading - page 3345

New comment
 
Aleksey Nikolayev #:

Thank you, a quality and interesting article with extensive literature.

It seems that they do not consider the kind of uncertainty that is interesting - probabilistic dependence of output on attributes. They study two other types of uncertainty - uncertainties related to inaccuracies of attributes and parameters. They are named beautifully - aleatoric and epistemic uncertainty) We should call our variant target uncertainty by analogy).

Imho, in our case "measurement errors" of attributes are absent in principle, and uncertainty of model parameters is poorly separable from our "target uncertainty".

It seemed to me that the sum of these uncertainties should give target uncertain ty. But I haven't really looked into it.

The approach is about the same as in kozula via meta lerners, but here we also propose a way to disassemble one model and use it as an ensemble of truncated classifiers, instead of an ensemble of several classifiers, for speed.

 
Maxim Dmitrievsky #:


I don't understand where the R square estimate comes from?

I was previously under the impression that this estimate is applicable in regressions if all regression coefficients are significant. Otherwise R squared does not exist....

 
СанСаныч Фоменко #:

I don't understand where the R square score came from?

I was previously under the impression that this estimate is applicable in regressions if all regression coefficients are significant. Otherwise R squared does not exist....

It's just something the tester shows for quick comparisons of different balance curves.

It's not involved anywhere else.

 
And it seems to me that the very direction is wrong in the root...
I think that it is necessary not to build the TS on all data, but on the contrary to choose one situation/pattern that already works At least 50/50 and try to separate works/not works, the usual binary classification.
 
mytarmailS #:
choose one situation/pattern that already works At least 50/50

They all work 50/50.

 
Ivan Butko #:

They all work 50/50.

It just seems like it.

It's like a 50/50 chance of meeting a dinosaur. It has nothing to do with actual probability.
 
mytarmailS #:
It just seems like--

It's like a 50/50 chance of meeting a dinosaur. It has nothing to do with actual probability.

If you score a figure in the script and look at the statistics of the future, the distribution of up/down, both by the number of candles and by the number of points tends to 50/50.

This is what concerns figures from candlesticks (the ratio of HLC with each other), and I did not count timeless ones, because they are too few for statistics of at least 1000 figures.

And so, if in 2022 the figure showed a forward in 55% of candles up and the average value of candles is 5-10% higher than in Sel, then in 2023 the payoff will still be 50/50, without any favours.

 
Ivan Butko #:

If you score a figure in the script and look at the statistics of the future, the distribution of up/down, both by the number of candles and by the number of points tends to 50/50.

This is the case with candlestick figures (the ratio of HLC with each other), and I did not count timeless ones, because they are too few for statistics of at least 1000 figures.

And so, if in 2022 the figure showed a forward in 55% of candles up and the average value of candles is 5-10% higher than in Sel, then in 2023 the working off will still be 50/50, without any privileges.

And if you add an adequate Stop and Take, will it be 50/50 too?

Or do you take profit and loss according to some definite average?
 
mytarmailS #:
And if you add an adequate Stop and Take, will it be 50/50 too?

Or do you take profit and loss according to some ephemeral average?
How can a statistical average be ephemeral?
It is what it is: on average so much up, on average so much down.
Based on them you can play with take and stop.

But this is a half-measure, because if TP and SL depend on the averages, they also work 50/50.

And if the averages are nothing to you, then TP and SL are pure fitting, pure 50/50, a toy in the optimiser.

The thought is different: statistics of simple patterns depends on long-term trends. And based on the work of manual traders, they trade independent patterns, which also in the long term down trade up in the plus.

But complex patterns rarely appear. The only option left is to ignore the small sample of statistics and try to feed them to the neural network.
 
To make an intermediate conclusion, both in kozula and variational and in all other directions, such as Bayesian classifier and generalised linear models, as applied to MO, ensembles are used.

In kozul, for some reason they are usually limited to one classifier or regressor (or two), aka meta lerner. Whereas in the paper about catbust and other libs, ensembles are used. Why exactly in kozula they don't reveal this is a bit strange. They don't generalise to the case of ensembles. It's basically just stats over models. No special magic there, but the results are sometimes pleasing.

Haven't seen any general reference book on the whole thing yet. It's kind of like ML.

And further on there is a fork in the road of how to apply all this to classification of time series, and to a special case - classification of BP for trading, the latter topic is practically not disclosed or mentioned anywhere.
1...333833393340334133423343334433453346334733483349335033513352...3656
New comment