weights= argument for froll{mean,sum,...}

[MichaelChirico]

Weighted means are pretty common in my experience.

The current workaround is pretty awkward (I might be using a weird approach, CMIIW):

x = 1:10
w = 10:1
frollapply(seq_along(x), n = 3, function(i) stats::weighted.mean(x[i], w[i]))
#  [1]       NA       NA 1.925926 2.916667 3.904762 4.888889 5.866667 6.833333 7.777778 8.666667

[jangorecki]

Little bit more explanation below.

verbose = function(i) {
  expr = substitute(stats::weighted.mean(.xi, .wi), list(.xi=x[i], .wi=w[i]))
  cat(i,": ", deparse1(expr),"\n")
  eval(expr)
}
frollapply(seq_along(x), n=3, verbose)
#1 2 3 :  stats::weighted.mean(1:3, 10:8) 
#2 3 4 :  stats::weighted.mean(2:4, 9:7) 
#3 4 5 :  stats::weighted.mean(3:5, 8:6) 
#4 5 6 :  stats::weighted.mean(4:6, 7:5) 
#5 6 7 :  stats::weighted.mean(5:7, 6:4) 
#6 7 8 :  stats::weighted.mean(6:8, 5:3) 
#7 8 9 :  stats::weighted.mean(7:9, 4:2) 
#8 9 10 :  stats::weighted.mean(8:10, 3:1) 
# [1]       NA       NA 1.925926 2.916667 3.904762 4.888889 5.866667 6.833333
# [9] 7.777778 8.666667

Weights has to be applied for each window separately therefore I don't see any shortcut here that could be applied. This needs to be done on low-level code.

[jangorecki]

There is a design decision what should be expected weights argument.

1. List of length same as length x (each rolling window can have own weights), each list element is numeric vector of length of value n

rollwm(x, n=3, w=list(rep(c(0.5, 0.3, 0.2), 5), rep(c(0.1, 0.2, 0.7), 5))

2. numeric vector of length same as length x

rollwm(1:10, n=3, w=10:1)

3. numeric vector of length same as value of n (all rolling windows uses same weights)

rollwm(1:10, n=3, w=c(0.5, 0.3, 0.2))

The first one feels like adaptive.weights, referring to currently implemented adaptive window sizes. Second and third can be achieved using the first.

@MichaelChirico example above asks for the second one. It is quite limiting as one cannot specify same weights for every rolling window.

Third one actually feels most reasonable and common case IMO, but of course limited to have same weights for every rolling window.

I would either go for implementation:

• only 3
• 3 and 1
• 3 and 2

WDYT?

[jangorecki]

After reading manual of stats::weighted.mean it feels much more clear that only 3rd implementation fits here.
Therefore this should be considered implementation to resolve this issue.
Implementing this on low level will be rather complex and might not happen anytime soon.