R - Fast way to calculate rolling mean with varying width

Question

I have a dataframe that contains bank assets for several dates (times). Each bank has a unique ID:

# Sample Data
time <- c(51, 52, 53, 55, 56, 51, 52, 51, 52, 53)
id <- c(1234, 1234, 1234, 1234, 1234, 2345, 2345, 3456, 3456, 3456)
name <- c("BANK A", "BANK A", "BANK A", "BANK A", "BANK A", "BANK B", "BANK B", "BANK C", 
          "BANK C", "BANK C")
assets <- c(5000, 6000, 4000, 7000, 8000, 10000, 12000, 30000, 35000, 40000)
df <- data.frame(time, id, name, assets)

> df
   time   id   name assets
1    51 1234 BANK A   5000
2    52 1234 BANK A   6000
3    53 1234 BANK A   4000
4    55 1234 BANK A   7000
5    56 1234 BANK A   8000
6    51 2345 BANK B  10000
7    52 2345 BANK B  12000
8    51 3456 BANK C  30000
9    52 3456 BANK C  35000
10   53 3456 BANK C  40000

For each bank I want to calculate the rolling mean of the assets, varying the width according to the number of consecutive time values. So the rolling mean shall include all availabe consecutive previous values of a bank's asssets. If there is no previous value availabe for one bank it shall equal assets. Therefor I add a column that counts the number of consecutive time-values and than use rollapplyrfrom the zoo package, which gives me the desired result, but with a large data set it is far too slow:

# Calculate number of consecutive times
require(dplyr)
df <- df %>%
  mutate(number.time = 1) %>% # insert column for number.time, start value = 1
  group_by(id) %>%
  arrange(time) # correct order for moving average

for(i in 2:nrow(df)) # Start loop in second row, end in last row of df
  df$number.time[i] <- 
    ifelse(df$time[i] == df$time[i-1]+1,    # Is time consecutive?
           df$number.time[i - 1] + 1,       # If yes: add 1 to previous number.time
           1)                               # If no: set number.time = 1
# Moving Average
require(zoo)
df %>%
  mutate(mov.average = rollapplyr(data = assets,
                                  width = number.time, # use number.time for width
                                  FUN = mean, 
                                  fill = NA,
                                  na.rm = TRUE))
Source: local data frame [10 x 6]
Groups: id [3]

    time    id   name assets number.time mov.average
   (dbl) (dbl) (fctr)  (dbl)       (dbl)       (dbl)
1     51  1234 BANK A   5000           1        5000
2     52  1234 BANK A   6000           2        5500
3     53  1234 BANK A   4000           3        5000
4     55  1234 BANK A   7000           1        7000
5     56  1234 BANK A   8000           2        7500
6     51  2345 BANK B  10000           1       10000
7     52  2345 BANK B  12000           2       11000
8     51  3456 BANK C  30000           1       30000
9     52  3456 BANK C  35000           2       32500
10    53  3456 BANK C  40000           3       35000

How could I get this output using a faster function? I'm aware of rollmean from zoo as well as SMA from TTR and mafrom forecast but these do not allow for varying width. My question may also be related to this question and this rblog, but I'm not familiar with C++ nor do I know a lot about function writing, so I do not really understand those posts.

EDIT 1: Note that in my code above it isn't the for-loop but the rollapplyr that takes a lot of time.

EDIT 2: The rolling mean shall include not more than the last 4 values. This is, as many consecutive values as there are according to the time-variable, but no more than the last 4 values. Sorry for the inexact question! :/ My wording was based on the assumption to use the "number.time"-column where it would have been easy to limit all values to maximum = 4.

Answer 1

First create a grouping variable g and then compute the rolling means. Note that rollsum is substantially faster than rollapply but does not support partial necessitating the workaround shown:

library(zoo) # rollsum

g <- with(df, cumsum(ave(time, id, FUN = function(x) c(1, diff(x) != 1))))
roll4 <- function(x) rollsum(c(0, 0, 0, x), 4) / pmin(4, seq_along(x)) 
transform(df, avg = ave(assets, g, FUN = roll4))

giving:

   time   id   name assets   avg
1    51 1234 BANK A   5000  5000
2    52 1234 BANK A   6000  5500
3    53 1234 BANK A   4000  5000
4    55 1234 BANK A   7000  7000
5    56 1234 BANK A   8000  7500
6    51 2345 BANK B  10000 10000
7    52 2345 BANK B  12000 11000
8    51 3456 BANK C  30000 30000
9    52 3456 BANK C  35000 32500
10   53 3456 BANK C  40000 35000

Answer 2

Use cumsum.

If you have just one bank, try:

cumsum(df$assets)/seq(nrow(df))

What to do if you have more than one bank, I leave as an excersize. Hint: you can completely avoid loops by using rle.

Here is the function "cumsum with restarts" which is supposed to help you.

cumsum.r <- function(vals, restart) {
    if (!is.vector(vals) || !is.vector(restart)) stop("expect vectors")
    if (length(vals) != length(restart)) stop("different length")
    # assume restart = FFTFFFTFFFFT
    len = length(vals) # 12
    restart[1]=T # TFTFFFTFFFFT
    ind = which(restart) # (1,3,7,12)
    ind = rep(ind, c(ind[-1],len+1)-ind) # 1,1,3,3,3,3,7,7,7,7,7,12
    vals.c = cumsum(vals)
    vals.c - vals.c[ind] + vals[ind]
}