Cross Validated is a question and answer site for people interested in statistics, machine learning, data analysis, data mining, and data visualization. It only takes a minute to sign up.
Sign up to join this community
I have the following problem of detecting the MA order in usage of the ACF plot.
Note: The process is stationary and the lags display monthly data.
I am not really able to say which order suits the data the best. In my opinion I could find argues for MA1 till MA12 and maybe even further.
With a stationary process, we can determine our $q$ parameter of the appropriate MA model by the number of significant lags in the ACF. The dotted-blue lines indicate the upper and lower limits for insignificant lags (based on the user-inputted confidence levels or default settings).
Some model comparison is ultimately necessary to figure out the best model due to differences in interpretation of the ACF lags. For instance, I would first compile an MA(14) model because I see the first 14 lags are each significant. Some discretion between using $q$ = 13 or 14 is necessary depending on how much you'd like to discriminate on significance. Many people choose to ignore the possible higher $q$ values suggested by higher lags that appear significant (look at the 21-27 range on your ACF plot), but it can't hurt to compare these models to the MA(14) for thoroughness. These may just be a product of some seasonality of your model.
In short: Start with choosing the highest significant lag in the initial grouping of your lags, then compare models on a test set.
