Machine learning in trading: theory, models, practice and algo-trading - page 2540

[JeeyCi]

JeeyCi #:

1) but maybe the "dynamic range" is painfully simple - the point of intersection of 2 MAs - it's important to get the periods right... Only OTFs look at 50 and 200... but for bigData analysis more profitable periods of MA may be found by the memory (weights) of neuronet (in case of other accompanying factors)... imho

it turns out that everything is simpler

MA is a kind of mechanical alignment

and you have to use analytical alignment (y-cap - aligned value in the model)... And it is better to use exponential dependence (since there are deceleration and acceleration stages in the trend) rather than power dependence (it only takes acceleration into account)... For nonlinear dependences, for example the analysis of series of dynamics (which can be only an additional method in adequate statistical research, but the analysis of dynamics never becomes the main method)

[JeeyCi]

JeeyCi #:

(I think I've seen somewhere recommendations for differentiation depending on 1s and 2s differences => choice of polynomial degree -- can't find it)...

something like this (from Shmoylova in "statistical theory")

[JeeyCi]

In general, in order not to jump out of the trend ahead of time, we should somehow conduct a full-fledged factor, correlation and regression analysis, and only then analyze the dynamics for acceleration, deceleration, reverse of the main trend... and do it somehow by sklearn, and only after that ML should be used to plot output on hyperplanes of bull/bear/hold-on... otherwise the commission will eat a lot from the depot... And I don't like 50/50 or even 25/50/25 probabilities... adequate money-management and risk-management...

stupid set of signs does not take into account interference

[JeeyCi]

And the most alarming thing in this whole story is that first we have to prove the normality of the NE distribution for a more or less representative statistical model... I can't prove it yet, so further evaluation is stalled... maybe, indeed, everything is not so random in the market, as Piligrim (his developer, I left the link above) thought

[transcendreamer]

Aleksey Nikolayev #:

Perhaps Eugene Fama in his dissertation, but I'm not sure.

Logarithmization is needed to make different periods comparable for sharply growing assets, for example bitcoin will have very different volatility in different years, which makes us invent and take some relative changes as a measure of volatility.

They also claim that logarithm mitigates heteroscedasticity and makes the distribution of regression model residuals more symmetric and slightly more normal, in practice everyone still screws it up... 😉

I agree, it's an unpleasant situation, because then I have to go back to the logarithm of the price, because broker does not allow to trade with logarithms, hehe...

[Aleksey Nikolayev]

transcendreamer #:

Logarithmization is needed to make different periods comparable for sharply growing assets, for example bitcoin will have very different volatility in different years, which makes us invent and take as a measure of volatility some relative changes.

They also claim that logarithm mitigates heteroscedasticity and makes the distribution of regression model residuals more symmetric and slightly more normal, in practice everyone still screws it up... 😉

I agree that in general it's an unpleasant situation because then you have to reverse the logarithm, because with logarithm prices the broker doesn't allow trading, hehe...

In my opinion, it's quite natural to use the logarithm - it's natural) Again, the intuition associated with interest should work - in fact, the rate of continuous accrual of interest is calculated (if you take the increment of the price logarithm and divide by time).

And different assets(in my opinion) are easiest to reduce to a common denominator by logarithmic price followed by normalizing by an average logarithmic spread.

[JeeyCi]

transcendreamer #:

It is also argued that logarithm mitigates heteroscedasticity...

I agree that in general it's an unpleasant situation because then you have to unlogarithmize

I'm not sure what the full extent of it is... only in the sense of asymmetry... but not in the sense of dispersion... imho

[JeeyCi]

Aleksey Nikolayev #:

I think it's natural to use the logarithm - it's natural.) Again, the intuition associated with interest is supposed to work - in fact, the continuous interest rate is calculated (if you take the increment of the logarithm of price and divide by time).

So intuition suggests that practitioners (not theorists) should take away the forward point to have a current price in futures (and there is no time in the price in the spot), while analysis of interest rates without time is indicative, when they are also floating (or are swapped to floating)... If we don't understand pricing of (derivative) assets, primitive mathematical transformations will only spoil the model... - An understanding of processes is essential in any modeling...

[mytarmailS]

Aleksey Nikolayev #:

The distinction between practice and theory works both ways. After practice usually just starts a new theory. Theory and practice are two legs that must be moved in turn and equally actively to get to the desired goal.

good words

[Mihail Marchukajtes]

mytarmailS #:

nice words

I agree, I support the same...