R: Loop requires long run time, suggestions on better structure - r

Question

I currently have a data set with quarterly returns for 10 indices. My dataset (compoundrates) is structured so that in the first column, we have "Scenario" and the second column is "Quarter", and the following 10 are the quarterly index return. The projection is 50 quarters, so lines 1-51 reflect quarters 0-50 for scenario 1, and lines 52-102 reflect quarter 0-50 for scenario 2, etc for 1000 scenarios.
To calculate cumulative compound rates, I need to multiply the current return by all previous returns from the projection. I set up a loop to do this in the code below:
for(i in 1:nrow(compoundrates)){
if(compoundrates[i, "Quarter"] == 0){
compoundrates[i, -c(1:2)] <- 1
} else{
compoundrates[i, -c(1:2)] <- compoundrates[i, -c(1:2)] * compoundrates[i - 1, -c(1:2)]
}
}
The loop is simple and works how I want. However, with 51000 rows, this takes about 13 minutes. Is there a way to speed up the code? I tried thinking of a vectorized solution, but could only think that I would need to loop through all rows of the dataset. While 13 minutes is not the end of the world, I have other datasets with longer projections, up to 200 quarters, which would take extremely long.
Possibly pivoting the dataset to be horizontal would require only 50 loops rather than 51000, but thought I'd see if anyone else had a more elegant solution.
Edit: Included here is a sample of the first couple of lines of my dataset:
> dput(head(compoundrates[, 1:4])) # First part of data, only 2 indices
structure(list(Scenario = c(1L, 1L, 1L, 1L, 1L, 1L), Quarter = c(0,
1, 2, 3, 4, 5), US = c(1, 1.06658609144463, 1.1022314574062,
1.1683883540847, 1.29134306037902, 1.28907212981088), MidCap = c(1,
1.10361590084936, 1.12966579678275, 1.21702573464001, 1.2674372889915,
1.37286942499386)), row.names = c(NA, -6L), class = c("grouped_df",
"tbl_df", "tbl", "data.frame"), groups = structure(list(Scenario = 1L,
.rows = list(1:6)), row.names = c(NA, -1L), class = c("tbl_df",
"tbl", "data.frame"), .drop = TRUE))

Answer 1

Try this out, it uses vectorized functions, basically exactly what you are trying to do with a for loop. It create new columns so you can see what is going on. Vectorized function usually run a lot faster than for loops
library(tidyverse)
compoundrates %>%
group_by(Scenario) %>%
arrange(Quarter) %>%
mutate(US_lag = lag(US),
MidCap_lag = lag(MidCap),
US_cum = US *US_lag,
MidCap_cum = MidCap *MidCap_lag) %>%
mutate_all(~ifelse(is.na(.), 1, .))
This should do the cumulative product you were asking for
compoundrates %>%
group_by(Scenario) %>%
arrange(Quarter) %>%
mutate(US_cum = cumprod(US),
MidCap_cum = cumprod(MidCap))
Per #carl-witthoft suggestion here is the benchmarking. I made a dataframe of 60,000 rows grouped by the 6 quarters in the OP.
Unit: milliseconds
expr
big_data %>%
group_by(Scenario) %>%
mutate(US_cum = cumprod(US),
MidCap_cum = cumprod(MidCap))
min lq mean median uq max neval
63.1487 70.0257 77.16906 73.72995 79.7645 147.167 100