Im trying to compare some anomaly detection methods for time series. Since there are not many time series with labeled outliers I decided to do a simulation study.

I'm doing the simulation in R using the "arima.sim" function which simulates time series from an ARIMA model.

Among the methods are some that are based on iterating through ARIMA processes (tso from tsoutliers package or X13-ARIMA-SEATS from Rjdemtra package) and other methods like those included in the anomalize package, that are based in decompose the time series in seasonality, trend and residuals and apply univariate anomaly detection in the residual.

My question is, is the simulation study fair? Are ARIMA error-based models not advantageous if the time series is simulated through an ARIMA process? Is there a method to perform a fair simulation for both methods?

Kind regards.

  • $\begingroup$ THE X13-ARIMA-SEATS model is a particular case of an ARIMA model as presented/discussed by W. S. Cleveland in a JASA paper , a few centuries ago.. AUTOBOX easily simulates an ARIMA model that the user wishes to specfy . stats.stackexchange.com/questions/461113/… might be of help. $\endgroup$
    – IrishStat
    Apr 24, 2020 at 17:08

1 Answer 1


I am not sure that "Since there are not many time series with labeled outliers" is true

The YAHOO datasets have over 200 real and synthetic labeled datasets. There are ten datasets here...https://www.cs.ucr.edu/~eamonn/discords/ etc

If you need more, email me eamonn keogh


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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