What makes an unsteady graph unsteady or why oil is oil? - page 25

 
Reshetov писал(а) >>

timbo, it has already been said that you need to learn the basics....


Yuri, if it's no secret, what are you really earning from, what's the hell of your system mix? I'm not asking to reveal the secrets of the system, I'm interested in the "direction of the road".
 
Reshetov писал(а) >>

The method of scientific reproduction is that those who don't believe, if they so desire, can double-check for themselves.


In your opinion, for predicting process x(i) = x(i-1) + e(i), e(i) ~N(0,1) will AR models work? Ready to do a demonstration experiment, let's have some fun ;)

 
Reshetov писал(а) >>

Learn your matchbook, timbo - it's tame.

Again, I would like a link to the math. There is a very famous ARPSS model and Box gives BP prediction within it. But it does not look as dashing as yours and has a lot of limitations.
 

Motivated by the message:

Reshetov писал(а) >>

Once again, for those who are particularly gifted:

1. from the initial BP we obtain the BP of the first differences

2. Extrapolate the BP of first differences

3. Reconstruct from the extrapolated part of first differences the extrapolated part for initial BP.


Experiment for the process x(i) = x(i-1) + e(i), e(i) ~N(0,1).

Plan:

  • generate data
  • cut the data into two parts (the first to build the AR model, the second to verify)
  • forecast the first part of the series and compare with the second

How we predict:

  • differentiate the series to be predicted
  • forecast by means of AR models
  • restore series from first differences

The course of the experiment:

  • Original data:
  • Two parts of the series:
  • Prediction result:
  • Fact/forecast separately:

It is not possible to make money using this prediction.

The reason for this prediction is the ACF of the process.

Conclusion: Despite the stationarity of the first differences the series turned out to be unpredictable.

Reshetov wrote >>

I'm not confused. If the first differences are stationary, the initial BP is predictable. This is quite obvious. If you hold a different opinion, then try to prove the contrary. And we will admire your efforts.


Your assumption is wrong, that was required to prove.

For the incredulous the mathcad file is in the attachment.

Files:
 
Reshetov >>:

Еще раз повторяю для особоодаренных:

1. Из исходного ВР получаем ВР первых разностей

2. Екстраполируем ВР первых разностей

3. Восстанавливаем из экстраполированного участка первых разностей экстраполированный участок для исходного ВР

You're not getting anywhere. You are confusing two things:

  • 1. Process model x(i) = x(i-1) + e(i), where e(i) ~N(0,1)
  • 2. Predictability of the process

Predicting a random process is only possible in the rms sense, i.e. recovering a "specific randomness" is simply useless (you have that in point two). But predicting the mean of a completely random process will not make you very romantic either. And after you find the theoretical/practical RMS error - everything will fall into place literally and figuratively. The prediction of an average will of course not be as "hopeless" as the pile of possible realizations, which for some reason did not happen, as shown by timbo. But there's no trading value to it.

(I've borrowed it without permission, but I hope timbo won't mind, I'm too lazy to make mine)


It is even more useless to apply such an elaborate approach to the market - the simple reason is that the distribution of incretions is completely different and leads to an even larger trajectory vector. But if you really want, there is a theory called "Asymptotic Random Walk Analysis". This theory quite thoroughly investigates deviations from initial conditions of the process (scientifically called "trajectory deviation"), including distributions of increments with heavy tails (including "different tails"). Used in risk theory, insurance and elsewhere.

ADDENDUM :

a little earlier, lea showed an example, and you can see that strangely enough, the first counts of the process are not that far off the average,


but still - you need something better for trading.

 

timbo, what did you draw this graph with, can I get a link to such a program?

 
Richie >>:

timbo, чем вы этот график нарисовали, можно ссылку на такую программу?

MATLAB. But the same can easily be done in any drawing program, even Excel.
 

It's been a while since I've been here. Stumbled across this thread. It's an interesting discussion.

First question to the participants: why is the first price difference stationary? Has anyone calculated the moments of such a process?

The second and more important question: why do some people believe that the stationary process is predictable? White noise is stationary too, but unpredictable. For those who do not believe, I can scientifically prove it. Or you can also do it this way. Imagine that white noise was predictable. Then the noise in the receivers wouldn't be a problem. Before receiving a signal we calibrate the receiver for external and internal noise and then at the moment of receiving a signal we begin to subtract the extrapolated noise from the noisy signal to get a clear signal. Shall we write a patent application together? :-)

 
Reshetov >>:

...

Доказательства я уже привел. Если Ваша ишачья упертость все еще не позволяет удостовериться в том...

There is a popular saying about people like you - "look in a book and see a fig". The only thing you have managed to prove, once again, is that you do not understand what you have read.
 
gpwr >>:

Давненько я тут не был. Вот наткнулся на эту ветку. Интересная дискуссия.

Greetings! Good to see.

First question to the participants: why is the first price difference stationary? Has anyone calculated the moments of such a process

For a very long time, some stationarity tests have been

Second and more important question: why do some people here think that a stationary process is predictable? White noise is stationary too, but unpredictable.

I don't know who, it's my first time here in a long time, but that's absolutely right - stationarity doesn't make the process predictable.

For those who don't believe me, I can scientifically prove it. You can also do it this way. Imagine that the white noise was predictable. Then the noise in the receivers would not be a problem. Before receiving a signal we calibrate the receiver for external and internal noise and then at the moment of receiving a signal we begin to subtract the extrapolated noise from the noisy signal to get a clear signal. Shall we write a patent application together? :-)

I've already figured out how to make money from a random process. :о) If I set parameters of the forecast to make balance become a stationary process (it will be a random one anyway) then I may earn some money (knowing parameters of initial process). If you save money for an extreme drawdown and as soon as you get the largest possible profit (RMS of the balance can be determined) you simply leave the market and do not show up again.)