I have a time series data set to which I want to fit an ARIMA model.

In looking at the plot of the data, it seems stationary, albeit perhaps marginally so. No trends, no seasonal effects, somewhat constant variance, etc. But despite the marginality, it seems as though it could quite possibly be stationary.

When I plot the ACF and PACF, I get the clearest possible signs that I should try an AR(1) model. The ACF is sinusoidal and gradually tapers to 0, and the PACF has one significant value that shuts off after lag 1.

When I fit an ARIMA(1, 0 ,0) model, the residuals plots are all just what I would want, and the ar1 coefficient is significant.

This seems like a good model, and so normally I would think I might want to consider it.

The problem is that an augmented Dickey-Fuller test indicates the series is not stationary, despite the plot of the series looking reasonable and the rest of the outputs looking good.

In light of the non-stationary indicator, I tried a first differencing.

Now the plot of the first differenced data seems to be unquestionably stationary, and the augmented Dickey-Fuller test is consistent with that conclusion. But when I plot the ACF and PACF of the differenced data, I get no significant values at all. This is the signature of white noise.

Nonetheless, I decided to fool around with it. Among other things I tried an ARIMA(1, 1, 1) model. Not only do the residual plots look good, and not only are the two coefficients significant, but the AIC, AICc, and BIC are all lower in this model than in the ARIMA(1, 0, 0) model.

I feel as though there’s something I’m really missing here. How can a series that has ACF and PACF plots indicating white noise actually be fit with a good model?

And given these two conflicting indicators, what would be the wisest choice?