I am new to ARIMA modeling and currently encountering a weird situation with time series of count data. The time plot shows clear seasonal patterns.ACF also hints on presence of seasonality. However, seasonal unit root test in R shows that series is seasonally stationary.

If I include seasonal differencing (D=1), I cannot find a single model where residuals satisfy normality assumption, even if I perform log- or - square root- or Box-Cox transformations of original series. If I do not include seasonal differencing (keeping seasonal AR and MA parameters is in the model), I easily identify a model with great diagnostics (residuals are white noise and normally distributed).

Having hard time solving the puzzle whether seasonal-looking data can be seasonally stationary. Will appreciate any suggestion.

  • $\begingroup$ Hello Jane, may I ask you which test you applied? Was it 'nsdiffs' from 'forecast' package? $\endgroup$ – DatamineR Jul 13 '13 at 10:32

Well, it's simply that clear seasonal correlations don't always imply that you need seasonal differencing; the seasonal effect can be modelled fine with seasonal moving average & autoregressive terms. It's the same in non-seasonal models; just because you see a strong autocorrelation doesn't necessarily mean you need to difference the series. Sounds like you solved the 'puzzle' by fitting a good model to the undifferenced series that still incorporated seasonal effects.

  • $\begingroup$ Many thanks Scortchi for your help! Now I am not worried about differencing. $\endgroup$ – Jane Apr 9 '13 at 14:13

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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