partial rolling sum in data.table in r

user_11740629 0
Viewed 583

I am looking for a fast way to implement a rolling sum on a big database. What I want is a fast function that supports both left (and right) alignment and an arguement for min observation.

Essentially, I want to calculate the rolling sum even if the sample size is smaller than the specified width. In the example below, I want the sum of next 5 values whenever next 5 values are available, and if the size of the leading values is smaller than 5, then summing the roll over all that is left.

Example: x <- seq(1:10)

Desired output:

15 20 25 30 35 40 34 27 19 10

I know that rollapply(x,5, sum, align = "left", partial=1) procudes the desired output, but am looking for a faster solution.

frollsum from the package data.table is fast but does not seem to have an arguement for min observation. roll_sum from the package roll is also fast and accepts a value for min observation, but does not support left alignment.

Thanks in advance.

rolling-computation data.table r
2 Answers

As of data.table_1.12.8, frollsum does not support align="left" with adaptive=TRUE, however you can tweak it as follows using rev:

library(data.table) #data.table_1.12.8

lrfrollsum <- function(x, k, align) {
    nk <- c(seq.int(k), rep(k, length(x) - k))
    switch(align, 
        left={ 
            rev(frollsum(rev(x), nk, align="right", adaptive=TRUE))
        },
        right={
            frollsum(x, nk, align="right", adaptive=TRUE)   
        })
}

x <- 1:10

lrfrollsum(x, 5, align="left")
# [1] 15 20 25 30 35 40 34 27 19 10

lrfrollsum(x, 5, align="right")
# [1]  1  3  6 10 15 20 25 30 35 40

Hopefully, this is fast enough.

user_19894800

I hadn't thought about using rev! Inspired by @chinsoon12 answer, another possible solution is

library(roll)
x <- seq(1:10)
rev(roll_sum(rev(x),5,min_obs = 1))
# [1] 15 20 25 30 35 40 34 27 19 10

Thanks!

user_117406290
Related