Question

I get why vectorised functions are better than for-loops.

But there are some problems where I can’t see the vectorised functional programming solution. One of those is summing monthly data to get quarterly data. Any suggestions to replace this code …

month <- 1:100
A422072L <- c(rep(NA, 4), rnorm(96, 100, 5) ) + 2 * month
A422070J <- c(NA, NA, rnorm(96, 100, 5), NA, NA) + 2 * month
Au.approvals <- data.frame(month=month, A422072L=A422072L, A422070J=A422070J)

Au.approvals$trend.sum.A422072L.qtr <- NA
Au.approvals$sa.sum.A422070J.qtr <- NA
for(i in seq_len(nrow(Au.approvals)))
{
    if(i < 3) next
    if(all(!is.na(Au.approvals$A422072L[(i-2):i])))
        Au.approvals$trend.sum.A422072L.qtr[i] <- sum(Au.approvals$A422072L[(i-2):i])
    if(all(!is.na(Au.approvals$A422070J[(i-2):i])))
        Au.approvals$sa.sum.A422070J.qtr[i]    <- sum(Au.approvals$A422070J[(i-2):i])
}

print(Au.approvals)

Now with enough data to run as an example.

Answer 1

Let’s create some bogus timeseries:

time_dat = data.frame(t = 1:100, value = runif(100))

To get a rolling sum, please take a look at rollapply from the zoo package:

require(zoo)
time_dat = transform(time_dat, 
                     roll_value = rollapply(value, 10, sum, fill = TRUE))

here I assume that the coarser resolution (quarterly) is 10 times coarser than the finer resolution.

---

Original answer for a non-rolling mean:

I like to use the functions from the plyr package, but ave, aggregate, and data.table are also good options. For large datasets, data.table is veeery fast. But to get back to some plyr magic:

First create an additional column which specifies the more coarse time frequency, i.e. which quarter is your observation in:

time_dat[["coarse_t"]] = rep(1:10, each = 10)
> head(time_dat)
  t     value coarse_t
1 1 0.9045097        1
2 2 0.4174182        1
3 3 0.5638139        1
4 4 0.8228698        1
5 5 0.7059027        1
6 6 0.5285386        1

Now we can aggregate time_dat for the coarser time frequency:

time_dat_coarse = ddply(time_dat, .(coarse_t), summarise, sum_value = sum(value))
> time_dat_coarse
   coarse_t sum_value
1         1  6.097348
2         2  4.834720
3         3  3.988809
4         4  4.170656
5         5  4.538269
6         6  6.198716
7         7  4.399282
8         8  5.507384
9         9  6.089072
10       10  4.663287

---