I have time series with daily data with, this series have 7 frequency. I used auto.arima in order to determine regressors. This function suggest me to use five regressor ar1,ar2,ar3,sar1 and sar2.You can see results below.


          ar1      ar2      ar3     sar1     sar2
      -0.7606  -0.4866  -0.1719  -0.6797  -0.3253
s.e.   0.0295   0.0344   0.0296   0.0284   0.0285

sigma^2 estimated as 2415:  log likelihood=-5929.68
AIC=11871.36   AICc=11871.44   BIC=11901.47

But actually my final goal is not to modeling with ARIMA. So I used arima only to find regressors and put them into some machine learning algorithm.

So my question is how to create or extract this five regressors ar1,ar2,ar3,sar1 and sar2 in separate time series?

  • $\begingroup$ The terms to which you refer aren't going to be time series, they are going to be numbers, as in the displayed printout. Or is it that you don't know what, for example, the ar1 value $-0.7606$ is a coefficient of? $\endgroup$ – jbowman Feb 23 at 23:45

These 5 impute lags of the output series GIVEN that the output series is regularly and seasonally differenced.

What you are looking for is an augmented data matrix of 0/1 values which can be pre-specified to put into a regression model.

Try identifying a sarmax model https://autobox.com/pdfs/PREFERRED.pdf and https://autobox.com/pdfs/SARMAX.pdf

Fundamentally your approach is flawed (arima) as you are implicitly starting with an assumption THAT all that is important is memory (no causals) . YOU need to focus on the DATA after it has been adjusted for causals and latent deterministic structure in order to identify the appropriate arima structure.

| cite | improve this answer | |

Not the answer you're looking for? Browse other questions tagged or ask your own question.