# Augmented Dickey Fuller says it's okay. But is it really?

This is original sales data. The biggest spike occurs during winter holidays with a consequent customer inactivity around early January. My Augmented Dickey Fuller Unit Root test is quite significant (t-test=7.227), which means that there is no unit root and the time series are stationary. What worries me is this huge spike and the sharp fall afterwards. Are we okay with using this data as long as ADF says it's okay? Or should we do something about it? Any thoughts are appreciated....

• ADF is not an omnibus test of all your model assumptions. It's not very sensitive to this kind of problem. – Glen_b Nov 26 '14 at 1:32
• Any suggestions? – Olga Nov 26 '14 at 1:33
• This is a very strange comment to make. Anyway, thanks for the input. – Olga Nov 26 '14 at 1:39
• I would try the dfgls test, which is more powerful, or maybe pperron. I would also consider adding the trend option to all three commands since you seem to have a trend on the right. More broadly, it would help with the interpretation if you also showed exactly what you typed (number of lags, trend, drift, constant). – Dimitriy V. Masterov Nov 26 '14 at 2:29
• This is where substantive knowledge of the data becomes important: what is the spike? – shadowtalker Nov 26 '14 at 7:27