[RsR] Time-series outlier problem

[Ajay Shah]

We've been using forecast::tsoutliers. In many situations it's nice, but I
have a MWE where it seems to get confused:

library(forecast)
x <- sin(seq(0,6.28,0.01)) # There's no noise

tsoutliers(ts(x)) # nothing is amiss.

x[4]  <- 2
tsoutliers(ts(x)) ## Weird

x[4]  <-  100
tsoutliers(ts(x)) ## Weird

It shows a lot of outliers and offers bizarre replacements.

We are working with time-series of about N=100 with gentle time series
structure. Do folks here have suggestions on how best to handle outlier
detection and correction for this?

-- 
Ajay Shah
ajayshah using mayin.org
http://www.mayin.org/ajayshah

	[[alternative HTML version deleted]]