#1
Hi Guys
I am doing time series regression
Had tried to find the order of correlation but the ACF function is a cyclical one (line a sine curve)
Then I tried finding the order of MA (Moving Average) process, which came out to be 7
However when I am charting residuals vs predicted value it is a downward sloping line
What do In make of it ?
 

noetsi

Fortran must die
#2
A MA of 7 is really high unless you have seasonality (and that would be an unusual lag to have seasonality on since it usually occurs at 12/3 or 4). Most ARMA models have 3 or less - in part because you can commonly model a high MA as a lower order AR process. Note that certain AR processes, negative autocorrelation I believe, show up as a sine wave.

I have not seen residual analysis such as you suggest for ARMA or ARIMA. It is more common to run a Box-Ljung test to see if there is residual autocorrelation after you have employed the filter. My guess is that you might have an AR process you did not account for (a mixed model) although I am relatively new to time series so take what I say with a large grain of salt:p
 

noetsi

Fortran must die
#3
A MA of 7 is really high unless you have seasonality (and that would be an unusual lag to have seasonality on since it usually occurs at 12/3 or 4). Most ARMA models have 3 or less - in part because you can commonly model a high MA as a lower order AR process. Note that certain AR processes, negative autocorrelation I believe, show up as a sine wave.

I have not seen residual analysis such as you suggest for ARMA or ARIMA. It is more common to run a Box-Ljung test to see if there is residual autocorrelation after you have employed the filter. My guess is that you might have an AR process you did not account for (a mixed model) although I am relatively new to time series so take what I say with a large grain of salt:p
 
#4
As the data has a weekly pattern, an MA of order 7 seems to satisfy it....But will check for a combination of MA and AR of lower orders.....I like your Yogi Berra quote too.
 

noetsi

Fortran must die
#5
The quote is half funny half serious as we generate numbers that involve a lot of money and impact on policy. Given the limits on time series, and our data, that makes me really nervous.

I have never dealt with weekly data. I guess a lag of 7 makes sense there. It would really be a SARIMA (that is seasonal arima) rather than ARIMA I would think although different software handles that in various ways. SAS has a method, based on theory developed several decades ago that uses IACF I believe to suggest mixed models. While less than ideal it can be useful. When I find it (I am collating my various sources these days) I will send you details if you are interested.

My favorite quote about ARIMA/ARMA/Box-Jenkins is that "...it is as much an art form as a science...." hardly a reasuring comment when important matters rely on your analysis.
 

noetsi

Fortran must die
#8
I don't think we can help with take home exams (and you would likely get in trouble if we did).

I don't have any problem with R. I am just giving Dason, Jake, trinker, spunky et el (who I talk to often in chat and are R fanatics) a hard time:p

Most of what I have I won't be able to send, but I will send some of the links as soon as I get it done.
 

noetsi

Fortran must die
#9
I don't think we can help with take home exams (and you would likely get in trouble if we did).

I don't have any problem with R. I am just giving Dason, Jake, trinker, spunky et el (who I talk to often in chat and are R fanatics) a hard time:p

Most of what I have I won't be able to send, but I will send some of the links as soon as I get it done.
 

noetsi

Fortran must die
#11
Thanks. I ran into a major problem today so it might be next week before I can send it. We just totally changed our process and none of the methods I know for time series will work in the future. (well until we have 48 months of data from the new process)...