[time-nuts] Outlier detection and removal

[Ole Petter Ronningen]

Hi, all

A reasonably well known method of outlier detection in phase data is to
convert to frequency and look for outliers more than k multiples of the
Median Absolute Deviation. I believe this is how outlier detection is done
in Stable32[1].

I have implemented an approximation to this method - I do not convert to
"frequency" as such, instead I simply take differences of subsequent phase
points. Then calculate the (absolute) median of the result, remove outliers
bigger than k multiples of the median, and integrate back to phase.

The results agree with Stable32 (bar a factor of 1.48-something on k) - the
same number of outliers are identified, so I have a reasonable confidence
in my approach. I have a gut feeling that my approach is equivalent to
converting to "proper frequency" - but I thought I would ask the more (than
me) mathematically gifted members of this lists if I am committing some
grave sin in my simplistic approach?

Thanks!
Ole

[1]
http://www.stable32.com/Outliers%20in%20Time%20and%20Frequency%20Measurements.pdf