[R-Forge #2187] Add/document rolling mean, median etc.. combined with i

[arunsrinivasan]

Submitted by: Matt Dowle; Assigned to: Nobody; R-Forge link

As highlighted in these :

http://stackoverflow.com/questions/1309263/rolling-median-algorithm-in-c
http://stackoverflow.com/questions/11676476/r-data-table-sliding-window

Functions to implement:

• rollmean
• rollmedian
• rollmin
• rollmax
  ...

[jangorecki]

can we have it included in 1.9.4 milestone?

also this link could be included as reference for FR: http://stackoverflow.com/questions/11676476/r-data-table-sliding-window

[jangorecki]

Couldn't that be done for "rolling function" instead? is it matter of optimization (for mean and median) behind the scene? if no, then I believe the documentation for custom user defined function would be much better because it could be easily applied on mean or median.

[arunsrinivasan]

Do you refer to functions having the name rollfun and then fun = "mean" etc instead of rollmean, rollmed etc? If so, I really haven't thought about it, but prefer to rollfun name with options inside (consistent with keeping related operations together in data.table).

[jangorecki]

Yes, exactly this is what I mean. Currently the most common name of such function would be rollapply from zoo package. There are also well developed lapply/mapply.
Here is the benchmark which includes all of the current approaches: performance-of-rolling-window-functions-in-r,

[eantonya]

Specifically for rolling mean/median/sd/min/max I think you need to beat or at least match caTools for this to be worth it.

@jangorecki arbitrary functions are going to be a lot slower than some specific ones: http://stackoverflow.com/q/18113323/817778

[arunsrinivasan]

@eantonya good to know. If that package already has optimised/fast functions, then maybe this needn't be a top priority then? We could try and implement as efficient as possible when there's more time, and then compare and export if necessary...

[jangorecki]

@eantonya caTools function runmean(x, k, ...) does not accept k as vector of corresponding window width for each observation. This is kind of hard limitation.
@arunsrinivasan In case if we would rely on external rolling function for now, it could be well documented how to use them most efficient with data.table. SO benchmarks can be also a good place.

[eantonya]

@jangorecki do you mean you want to have a separate k for each group (in which case this is very easy to do with runmean) or do you mean smth else?

[arunsrinivasan]

@arunsrinivasan In case if we would rely on external rolling function for now, it could be well documented how to use them most efficient with data.table. SO benchmarks can be also a good place.

Yes that makes sense. But if you could clarify your earlier point about “does not accept k as vector” (maybe with an example), it’d help prioritise.

[jangorecki]

on the common zoo usage:

library(zoo)
x <- runif(1000,0.5,1.5)
width <- rep(seq(from = 5, to = 20, by = 5), length(x)/4)
str(width)
# SMA period weighted by custom width
SMAvar <- rollapplyr(data = x, width = width, FUN = mean, fill = NA)

in the mentioned SO performance-of-rolling-window-functions-in-r there are four approaches to achieve that - one is simple for loop example so also can be helpful.

[arunsrinivasan]

@jangorecki 👍 thanks!

[jangorecki]

Leaving for reference another example. This time rolling sum by group for irregular time series: http://stackoverflow.com/a/27983553/2490497
Thanks to shift.