First I'd like to say thanks a LOT for this bringing Rodionovs method to R, @alexhroom ! Without your work it would still be on the brink of oblivion due to the unmaintained excel macro before Then, I agree with @jatalah that it would be great if Huber's weight function could be added. Without the Huber function, the Rodionov's method is really sensitive to outliers as they are often found in real-world data.

Many thanks

@jatalah @tjgrabs sorry for little progress on this so far - most explanations of how Huber's weight is used is a little less than clear, and the work is being held up by me looking for a better & more mathematical explanation in the literature (e.g. Rodionov's description says that the outliers are handled using Huber's weight, but doesn't particularly explains where this fits into his original 2004 algorithm). furthermore, as I do not have Microsoft Office my access to his excel macro is limited. if anyone can provide a paper describing the algorithm with huber's weight added it would be a much faster task, otherwise it might take a while as I need to find the time to better digest the use of the parameter.

Dear @alexhroom,
I just discovered that Sergej has a website with a more detailed description of the different algorithms: https://sites.google.com/view/regime-shift-test/home. The site also contains links to different implementations, including the rshift package and a new, updated excel version. I attached extracted code from the excel version, which you might also use to see how the huber weighting was implemented.

ExcalMacroModule_Code_BaseShiftInMean.txt

Apart from this, I tested rshifts on a data set, and noticed two things: 1. The results differ from Rodionovs Excel implementation, even after I use a large Huber parameter (which I think should result in ignoring outliers). 2. Two breakpoints appeared in adjacent years. This is not a behavior I would expect. The results from the Excel plugin shows only a single breakpoint in 1973 and not in 2000. (see attached image).

@tjgrabs thank you for this link. the difference in results is a known discrepancy - there’s some minutiae in the original Rodionov paper that allows for room of interpretation, which can cause different implementations to have slightly different sensitivities - it should only affect sensitivity of the algorithm though and shouldn’t be hypothesis-affecting for research. I’m rewriting my implementation to be faster and more modifiable which may fix this - can you please share with me the data set you tested on?

EDIT: This issue should be fixed with rshift v2.20, which uses a Rust implementation that as far as I can tell gives results equivalent to Rodionov's Excel version.

There is a GitHub repo for the Stirnimann 2019 paper that has code for the STARS regime shift test, weighted averages using Huber's weight parameter, and the various pre whitening routines. The last commit was from 5 years ago, but the foundational pieces are there.

@adamkemberling I'd been working on the Huber's weight things in issue #24. The main snag I ran into was being unable to find some good test data to check whether I'd done it correctly, and then I didn't have any time to work on it for a while. I don't have access to Microsoft Excel so I can't get some results from the Excel macro for comparison!