0
$\begingroup$

What's the best way to find outliers in the time-series like the one attached?

Just using the knowledge that it is a measurement of a continuous physical process. The data is quite short (200-500 samples) and it is not really periodic. enter image description here

Thanks

$\endgroup$
1
  • $\begingroup$ Try this quick intuitive scheme. Proceed to fit linear segments using Least-squares. Points that are outliers may then be identified as departing from the confidence interval (CI) as established for the local regression segment. Trial and error may establish a working optimal linear segment length (or # of moving points) and CI band-width for the local regression segments. Concept: The piecewise regression supplies an estimate of the local mean adjusted for trend, and a variance estimate for outlier decision making. $\endgroup$ – AJKOER Oct 6 '20 at 18:50