Online mean shift algorithms I am looking for an online algorithm which can identify mean shifting in a time series quickly, I have seen some algorithms that do so but they require 50+ samples in order to flag that the mean has changed
any suggestions?
Thanks
 A: AUTOBOX http://www.autobox.com/cms/index.php/afs-university/intro-to-forecasting/doc_download/53-capabilities-presentation can be very useful for this. As @EdM wisely suggested one needs to be able to control for false positives i.e. level of confidence BUT also be able to incorporate/specify a   minimum magnitude of the identified/proposed level shift. This is a very important and unique (to my knowledge) option when using AUTOBOX a program that I have helped develop. Their 30 day free version could be useful to you as a possible springboard to help you solve your (very common but often unstated) problem. An interesting feature is the capability to identify/isolate local time trends which often are falsely reported as level shifts. Hope this helps.
There is no hard and fast rule for a minimum sample size as the important idea is not necessarily the number of observations but the signal to noise ratio for the effect versus the error process. The more pronounced/larger the ratio the smaller the number of observations that are needed to "identify/capture" the effect. I have examined many of the earlier attempts to perform this task and find that collectively their greatest failure is failing to seamlessly/correctly incorporate memory and remedies for various alternative Gaussian Violations into the heuristic mix.
