What is the go to algorithm for finding an anomalous discord (sequence) in a time series given only the time series itself rather than being able to compare to other time series?
(I have access to Python and R.)
$\begingroup$
$\endgroup$
1
-
$\begingroup$ I would refer to the paper by Tsay in 1989. $\endgroup$– Tom ReillyCommented Oct 5, 2017 at 17:16
Add a comment
|
1 Answer
$\begingroup$
$\endgroup$
tsoutlier can be a reasonable attempt to detect anomalies. Unfortunately it assumes that an ARIMA model be prespecified prior to detection and does not work when stochastic causals are in play. In other words memory is more important / dominant. To identify an ARIMA model using freely available solutions requires that there are no outliers, constant error process and constant parameters over time which is flawed if there are anomalies present. Thus one needs to simultaneously identify anomalies and ARIMA model form (along with variance / parameter stability remedies) via a more comprehensive search process.
