Not Mashka's business! - page 2

New comment
 
Neutron:
lna01:
I would look at how the weights behave on history. That is, I would make an indicator with three buffers: w1,w2 and w3.

No problem. Only what will it give us? It is clear that they will behave regularly with a period of fluctuations of a smaller scale, because they are a solution of the cubic equation.

I would try to get a visual impression of their predictability and possibility of existence of some patterns from such a graph.


P.S. If the impression was negative, I would probably refuse to continue.

 

Seryoga, if you don't understand this: "If you reconstruct the time series by the predicted average, it won't work, big errors", then it's simple. I meant that if you have predicted the MA curve for N counts ahead, and knowing the initial BP, you can easily reconstruct the future BP.

 

5+ to Neutron

The method of weeding out "line-dependents" is very good in itself.
Otherwise some people take it from the ceiling, or do mysticism 5, 8, 13.

 
Yes, curious about the 6, 80, 300. Maybe that's the twist on Better's system - given that, say, on the minutes, 80 is 5.33 on the 15 minutes (oops, non-integer periods appeared), and 300 is about 5 on the hours?
 
By the way, yes, it is also possible to reverse the problem. To see how the periods of the flapping should change, so that the weights remain approximately constant. The correlation threshold of 20%, judging from the first post, is not really justified.
 
Ahh, I see who the perfect ma is. I should have drank less ale last night :o)
 
lna01 писал (а): The correlation threshold of 20%, judging from the first post, is not particularly justified.

It's actually not a bad threshold, not the nasty 60-70% you can't do much good with. I tried once, when I was messing around with NS, to see how the forecast's S.T.O. varies depending on the correlation of several forecast series. The conclusion was that if there is correlation and it is positive, there is a limit to the decrease of the OR, i.e. it does not decrease at all inversely proportional to the root of the number of series.

 
Mathemat:
lna01 wrote (a): The correlation threshold of 20%, judging from the first post, is not particularly reasonable.

It's actually not a bad threshold, not the nasty 60-70% you can't do much good with. I tried once, when I was messing around with NS, to see how the S.O.P. of the forecast changes depending on the correlation of several forecast series. The conclusion was that if there is a correlation and it is positive, there is a limit to how much the correlation decreases, i.e. it does not fall at all by the inverse of the root of the number of rows.

This is not so smooth. The length of the row is very important for calculating the AC, and in fact it is tantamount to a ceiling from which to take values :o(

 
lna01:
Neutron:
lna01:
I would look at how the weights behave on history. That is, I would make an indicator with three buffers: w1,w2 and w3.

No problem. Only what will it give us? It is clear that they will behave regularly with a period of fluctuations of a smaller scale, because they are a solution of the cubic equation.

I would try to make a visual impression about their predictability and possibility of the existence of some patterns.


P.S. If the impression was negative - probably would have refused to continue.

Aha, I get it! Indeed, if the characteristic chattering period of the coefficients is less than the averaging window N, then you can forget about prediction. And that's exactly what will happen. Thanks, Candid, you've just saved me a lot of time and effort. I can see that the problem can't be solved in this formulation.


grasn 10.04.2008 14:19

Seryoga, if you don't understand this: "If you reconstruct the time series by the predicted average, nothing will come out, big errors", then it's simple. I meant that if you have predicted the MA curve for N counts ahead and know the initial BP, you can easily restore the future BP.


That's the thing, I couldn't find a way to "easily recover" the original BP. All the methods I know fall apart when approaching the right edge of the BP. I even once posted a cartoon on this forum where the process of approaching the forecast series to the event horizon is shown. The matter is that by integrating initial BP (constructing MA) we practically do not bring anything new to processed data and therefore we don't get any progress in terms of forecasting. I think we need a tool capable of analyzing non-linear BP dependencies...


Korey 10.04.2008 14:26

5+ to Neutron

Thank you!


 

to Neutron

Seryoga, I'm a bit confused (don't pay attention. It's residual from beer :) Please, how did you calculate the mutual correlation between MA? [MA(n) and MA(n+1)] then[MA(n+1) and MA(n+2)] or in some other way?


If so, and observing the trend of the graph itself:


It is not quite clear where these values come from. After all, starting from a window of length 20 and above, the correlation between MAs is very strong and how they differ by 20% and then how you got windows 6, 80 and 300. This is hardly possible! But if you calculated e.g.[MA(n) and MA(n+k)], then on what basis did you choose this k (thinning conditions)? Does choosing k change the result?

В том-то и дело, что я не смог найти способа "легко восстановить" исходный ВР. Все известные мне методы рассыпаются при приближении к правому краю ВР. Я даже как-то мультяшку выкладывал на этом форуме где показан процесс приближения прогнозного ряда к горизонту событий. Дело в том, что интегрируя исходный ВР (строя МА) мы по сути ничего нового не привносим в обрабатываемые данные и, как следствие, не продвигаемся в плане прогнозирования. Думаю, тут нужен инструмент способный к анализу нелинейных зависимостей ВР...

OK. I'll give you my humble idea later :o)

123456789...12
New comment