Remembering veterans: Box and Jenkins - page 3
You are missing trading opportunities:
- Free trading apps
- Over 8,000 signals for copying
- Economic news for exploring financial markets
Registration
Log in
You agree to website policy and terms of use
If you do not have an account, please register
Had a look at Yudin. Not impressed. Tutorial. Little part of EViews or STATISTICS. Debugging of the programs is unknown. The tastiest bits are missing. Like the first tutorial. But nothing industrially applicable.
I didn't buy this textbook for impressions. For the applied tasks that I have to solve, the programmes described in it are over the top, i.e. I don't use much of the functionality. Obviously, the tutorial doesn't contain examples for all occasions, and it's just for understanding the basics, and then on your own.
As for the debugging of software, it has some bugs, but IMHO in quantity and quality, not beyond the limits of proprietary analogue software.
There is an extension of the ARSS model in the form of ARPSS, where P is pro-integrated. This is where it comes in. Integrated means differentiated! That is, the difference between the neighbouring bars of the quotient is taken!
In this case, "P" stands for "Read the theory.
"The series y[t] is called integrated because it is the result of applying to the stationary series w[t]=(1-L)^d*y[t] the operation of accumulated sum d times" // http://quantile.ru/01/01-AT.pdf section 3.4 "ARIMA process prediction"
(3) Completely new result for TA: coefficients in an indicator are random variables. At least one conclusion: indicators without adaptation of coefficients to the current quote are meaningless.
You forgot to say - when building the model you implicitly assume that the coefficients of the model generating the process are constant. But since you don't know them - you resort to estimating these coefficients based on the observed realization of the process. And in this case your estimates are random variables.
If you believe that the coefficients of the model generating the process are random variables - you must use either other models or adequate methods of parameter estimation.
So the conclusion you draw is completely wrong.
I didn't buy this textbook for impressions. For the applied tasks that I have to solve, the programmes described in it are over the top, i.e. I don't use much of the functionality. Obviously, the tutorial doesn't contain examples for all occasions, and it's just for understanding the basics, and then on your own.
As for debugging software, so it has some bugs, but IMHO in quantity and quality, not beyond the limits of proprietary analogue software.
I think there are many in Excel.
By the way, haven't seen a unit root test.
Only in responding to your post did I realize that a forward TC test (including NS) only makes sense if the quotient is stationary. First the unit root test and then the forward test.
didn't understand a word =)
And it doesn't hurt, because everything is in Russian about events of long ago.
In this case the "P" stands for "Read the theory".
"The series y[t] is called integrated because it is the result of applying the cumulative sum operation d times to the stationary series w[t]=(1-L)^d*y[t]" // http://quantile.ru/01/01-AT.pdf section 3.4 "ARIMA process prediction"
You forgot to say - when building the model you implicitly assume that the coefficients of the model generating the process are constant. But since you don't know them - you resort to estimating these coefficients based on the observed realization of the process. And in this case your estimates are random variables.
If you believe that the coefficients of the model generating the process are random variables - you must use either other models or adequate methods of parameter estimation.
So your conclusion is completely wrong.
Bah, colleague, where were you before. Join us.
If you believe that the coefficients of the model generating the process are random variables - you must use either other models or adequate methods of parameter estimation.
So the conclusion you have drawn is completely wrong.
Strictly speaking, you are right.
But I am pushing the idea that those coefficients which we see as a result of estimation are not constants, but estimates lying within certain intervals. It is obligatory to pay attention to the following columns of the table for coefficients.
I do not see how non-stationarity and memory can be equated.
No, of course. quite different concepts and they are treated as separate, but within the same approach.
Non-stationarity is defined by the unit root test.
There is no concept of "memory" in the classical work. But the concept of ACF and CHAKF is widely used. It is suggested to determine the order of ARIMA model by the appearance of these graphs. But it is possible to do something simpler: to find the minimum model by a search method.
Yeah, and in econometrics the term "memory" has been replaced entirely by "ACF". Wonderful.
Don't. I won't quote the fable.
Only discussing Box and Jenkins with some extensions. With them ACF plays a very important role. I pointed out which one.
Speaking of the value of the thread.
I have posted many times about the caution of interpreting test results and forward tests. Only today in response to Reshetov I have realized one very important idea. Testing and forward testing can only be trusted if the residue = kotir - TC is stationary! i.e. passes the unit root test. If this residue is non-stationary, then testing cannot be trusted at all, and no forward tests will help. It seems to me that a branch could have been started for the sake of such a conclusion. Here's Box and Jenkins for you.