I ran various diagnostic tests on a time series dataset using R. The null hypothesis for non-stationarity was not rejected using the Dickey-Fuller test, and moreover the null hypothesis for the Ljung-Box test was not rejected either; implying residuals follow a random pattern.

When I run the acf and pacf plots, the acf gradually decreases while pacf cuts off sharply after the first lag. However, when I run auto.arima the model is specified as (0, 1, 0) - random walk with drift.

Am I right in saying that the ARIMA calculation conflicts the findings of acf/pacf, since the acf/pacf result implies an AR(1) process?

I would appreciate any advice as to how I should proceed from here, or what weight I should give to the various results in my interpretation.

  • $\begingroup$ Post your data. $\endgroup$
    – Tom Reilly
    Commented Mar 31, 2016 at 18:09

1 Answer 1

  • ACF/PACF indicate AR(1)
  • Dickey-Fuller indicates a unit root

Hence, both taken together indicate an AR(1) with a unit root, another name for which is a random walk.

  • auto.arima selects random walk with drift

This is quite in line with the above except perhaps for the drift. However, depending on the specification of the (presumably augmented) Dickey-Fuller test, you might have a drift there. In sum, there is no clear evidence of conflict between {ACF/PACF and Dieky-Fuller} and the model selected by auto.arima.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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