Machine learning in trading: theory, models, practice and algo-trading - page 2764

[СанСаныч Фоменко]

Evgeni Gavrilovi #:
After oversampling, will the new predictive power of the predictors be of value? Who has thought about that?

It doesn't. Checked.

[Ivan Butko]

Just a curious fact: the neural network determined the importance of mashka with a shorter period than with a longer one))) It seems to be a discrepancy with the common understanding: traders usually say that "heavy" mashka is more important, the price beats on it with its horns, if it is a bull, and crushes it with its paws, if it is a bear.

Below, mashka with periods 512-256-128-64-32-16-16-8-4-4-2-1 and their importance are fed to the input (OUT here is also an input).

[mytarmailS]

Ivan Butko #:
Just a curious fact: the neural network determined the importance of mashas with a shorter period than with a longer period.)) Kind of a discrepancy, ......

It got it right...

Slow process as a result of fast process, but not vice versa...

You can predict tics by tics, but you can't predict tics by tics.

So the faster Mashka is, the better in this situation.

And the fact that mashka with a long period is important for traders, it's different, a rebound from a very slow mashka is a rare event, say once in 500 bars, and you probably predict every new bar, do you understand the difference?

So the neuron doesn't lie, the problem is experimentation.

[Ivan Butko]

mytarmailS #:

Slow fly, it's a rare event, say once every 500 bars, and you're predicting every new bar?

According to my observations, if you add to the neural network something "axe", "heavy", but related to the instrument, then on the one hand it distorts the balance line on the backtest after training. BUT! It also equalises the forward line, it stops going to the bottom every time and goes flat. That is, roughly speaking, we won't make money, but we won't lose the whole depo, and this is the first rule of trading. It turns out that we should juggle in this direction, because the next task is to turn the balance line upwards.

[Valeriy Yastremskiy]

Maxim Dmitrievsky #:

I don't know what you're talking about.

I don't use time, it's built into the increments.

Last time the trading time was specific, I don't remember by day of the week. but definitely by time of day, from 17 to 18 like. I.e. manual time breakdown or something else, but final training on a specific time period.

In general, the construct If increments are such and such in such and such a time interval is used without splitting into time segments or first find time segments that have something on them and then train on them?

[Valeriy Yastremskiy]

СанСаныч Фоменко #:

Time-based signs are not used. They have low predictive power and very large variations in that predictive power. Although they are better than the mashka variants.

Thanks.

[Maxim Dmitrievsky]

Valeriy Yastremskiy #:

Last time the trading time was specific, I don't remember the days of the week. but exactly the time of the day, from 17 to 18 like. I.e. manual or otherwise, but final training on a specific time period.

In general, the construct If increments are such and such in such and such a time interval is used without splitting into time segments or first find time segments that have something on them and then train on them?

Can I have a link? I don't know what we're talking about.

[JeeyCi]

Evgeni Gavrilovi #:
After oversampling, will the new predictive power of the predictors be of value? Who has thought about this?

SanSanych Fomenko #:

It doesn't change. Checked.

If a person doesn't distinguish between covariation and correlation, I don't even ask such a person what they mean by median

SanSanych Fomenko #:

, but "predictive ability", understood as the difference between the medians between two vectors obtained by dividing the predictor by classes, is absolutely accurate.

By description, this is the "line" which in ML is just a threshold in any classification algorithm....

if he had done a standard intergroup analysis of variance, he would have been able to estimate statistical significance, but of course oversampling doesn't change anything for him (it just counts %% of correct guesses of class membership)....

after his reference to the picture of covariances, I can clearly state that he is comparing flies with cutlets... which proves this question of his (he forgets and remembers correlations very slippery).

SanSanych Fomenko #:

1. Do you know the correlations between actual and nominal variables?

I know OLS and ANOVA and the interpretation of the significance of their estimates, and the fact that resampling doesn't change anything for your "ability" can only indicate that the "if-then" function would be enough for you not to build a model (and even try to ignore the statistical basis of modelling) and not to be able to estimate the coefficient of significance of the results of your classDist, but only the percentage of reliable answers given by it....

== the same problem of indistinguishable flies and cutlets in basic concepts with cries about "tools" that hit on or off the mark.... well, yes, with a 30% probability (some tools are still slightly distinguishable) - maybe "do nothing" is distinguishable by 70% guessing ?

It just can't get the change of coefficients_values during resampling from its algorithm... by definition

[Valeriy Yastremskiy]

Maxim Dmitrievsky #:

Can I get a link? I'm not buying it. What are you talking about?

I can't find the correspondence here, I have the file here, July 2020.

[Maxim Dmitrievsky]

Valeriy Yastremskiy #:

I can't find the correspondence here, I have the file here, July 2020.

I don't remember what it is ) maybe it's from an article on season tickets.

in any case, I don't touch time now, if the analogy with the stumbling window, it's not about time as such.

but in general yes, the lag of increments will change and the window itself will shift. It can be done in different ways, I haven't done something specific yet.