Machine learning in trading: theory, models, practice and algo-trading - page 3707

 

rates are still "conditionally constant"... currency A to B rises/falls at a rate of X%/day. A zigzag with a constant angle of inclination. The duration of these segments is poorly predictable and the change of movements due to volatility can be seen afterwards, when it is too late to throw oneself away

 

It is enough to choose one mode, optimise the TS on it and compare it with the optimised TS on all data, on the forward. So as not to drown in holography. There are articles on this topic.

It is possible to search for modes in the optimiser. The advantage is that you don't need to optimise the ranges of features in the search for modes, they are already there. There will be no fitting of ranges to other variables when it turns out that it is not a real mode, but just a fitting of a number of values to other values in the TS.

Some kind of conditional-hypothetical grails are off the table.

 
Maxim Dmitrievsky #:
Well, for example, volatility is conditionally constant.

I agree. Either it is almost constant, or it changes (on minutes, for example) according to a constant periodic law.

It is not clear only how this constancy can bring profit. Roughly speaking, SB also has constant volatility.

 
Maxim Dmitrievsky grails are off the table.

Of course testing in practice is more important than any theory. If it takes off, it doesn't matter so much why).

 
Maxim Kuznetsov #:

rates are still "conditionally constant"... currency A to B rises/falls at a rate of X%/day. A zigzag with a constant angle of inclination. The duration of these segments is poorly predictable and the change of movements due to volatility can be seen afterwards, when it is too late to throw oneself away

I guess the statistics will not show much difference from the SB. Imho, in zigzags the difference from the SB should be found in the way larger zigzags are formed from smaller ones (Mandelbrot and inherent fractals).

 
Aleksey Nikolayev #:

I suppose the statistics will not show much difference from the SB. Imho, in zigzags, the difference from the SB should be found in the way larger zigzags are formed from smaller ones (Mandelbrot and its inherent fractals).

By statistics they (const slopes) and are detected against the background of other things, somehow no longer SB. And they can be calculated analytically, because they are a consequence of "conditionally constant" volatility, i.e. they are not SB at all

The slope is not a random variable. (it can be assumed that it is a "convenient, habitual" value for exporters/importers. But this is more fundamental analysis, not games with packs of quotes).

But how to catch in real time the moments of reversals or to determine the duration of the knees of such a zigzag, that is the question :-)

 
Maxim Kuznetsov #:

They (const slopes) are detected by statistics against the background of other things, somehow they are no longer SB. And they can be calculated analytically, because it is a consequence of "conditionally constant" volatility, i.e. not SB at all.

The slope is not a random variable. (it can be assumed that it is a "convenient, habitual" value for exporters/importers. But this is more fundamental analysis, not games with packs of quotes).

But how to catch in real time the moments of reversals or determine the duration of the knees of such a zigzag, that is the question :-)

Nothing is clear, but very interesting. Some set of assertions follows from what is unclear. However, as usual.
 
Aleksey Nikolayev #:
I don't understand anything, but it's very interesting. Some set of assertions that follow from nothing. As usual, though.

I have already told years ago and repeatedly how and what counts. Everything is simple, but everyone is fascinated by new ML toys, now AI too :-) Nobody knows how to preprocess and really analyse data, and nobody wants to.

Everyone wants to throw arbitrary methods into a pot, pour data into it and the pot should somehow boil itself and give them the grail. But it gives out crap and everyone is surprised, but they are measuring varieties and consistency. Good luck to you

 
Maxim Kuznetsov #:

I have already told years ago and repeatedly how and what counts there. Everything is simple, but everyone is fascinated by new ML toys, now AI too :-) No one knows how to do preprocessing and real data analysis, and no one wants to.

Everyone wants to throw arbitrary methods into a pot, pour data into it and the pot should somehow boil itself and give them the grail. But it gives out crap and everyone is surprised, but they are measuring varieties and consistency. Good luck to you

Well, then you would have provided links to "told earlier".

So far it looks more like "told before" from Renat. By the way, I haven't seen him on the forum for a long time, probably busy packing his trillions of money.

 
Aleksey Nikolayev #:

I agree. Either it is almost constant, or it changes (on minutes, for example) according to a constant periodic law.

It is not clear only how this constancy can bring profit. Roughly speaking, the SB also has constant volatility.

Well, further I think from the type of strategy. For example, if it is a mean reversion - some signs and modes are suitable for it, trending - others. In principle, there are not so many signs hidden in a one-dimensional series of quotes that you can miss :). And then - as luck would have it.

Suppose, according to Mandelbrot, if H != 0.5, then there are some inefficiencies, which theoretically can be transformed into a TS. But H is variable, and in a particular regime it may turn out to be constant. That is, so that at least something is constant. If multifractal spectra are calculated, there will be several components with different H.

It is difficult to say in some normal words what exactly is being searched for :)