I've been working on the time series prediction of a signal and came across a small misunderstanding. The signal is depicted below: enter image description here Apparently it looks like there are several stationary local areas but overall it does not look stationary to me. However, ADF confirms the stationarity with the p-value to be almost 0.

ADF Statistic: -6.554090 p-value: 0.000000 Critical Values: 1%: -3.431 5%: -2.862 10%: -2.567

When I draw a graph of first differences, it visually looks stationary: First differences

ACF of the original signal does not look promising:

I am now confused in the approach to predict, should I use the signal itself or first differences to build an ARIMA model? What would be your approach?



closed as unclear what you're asking by Michael Chernick, Siong Thye Goh, BruceET, Jeremy Miles, jpmuc Oct 21 at 18:25

Please clarify your specific problem or add additional details to highlight exactly what you need. As it's currently written, it’s hard to tell exactly what you're asking. See the How to Ask page for help clarifying this question. If this question can be reworded to fit the rules in the help center, please edit the question.


ADF test is not test for stationary.
ADF test is for if data is unit root or not.

Your data seems to have several change point, I recommend to separate dataset.
Then check adf, auto correlation, seasonality.. again.
If it has enough auto correlation, you can use ARIMA model.

  • $\begingroup$ I am not well versed in topic, but what is the idea behind separating time series? $\endgroup$ – Oleksandr Puhachov Oct 16 at 18:27
  • $\begingroup$ For example, in observing motor temperature, stopping motor and accelerating motor have completely different physical phenomenon. machine learning is hard to describe those at the same time. So separate dataset by the situation in especially univariate data. $\endgroup$ – Nokius Oct 17 at 0:52
  • $\begingroup$ How one would then make a prediction when the new data comes? The graph above is a CPU voltage which changes based on a server workload (I should have perhaps mentioned this in the question). The server is running continuously (except the maintenance days, but those are not in this time series). Workload basically comes on demand (let's say, someone needs to run an application on server). I understand changes perhaps can't be predicted due to a random nature, but I am looking to capture a trend (let's say, change was predicted with a delay, but then prediction follows the trend) $\endgroup$ – Oleksandr Puhachov Oct 31 at 14:45

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