Rolling sum with a conditioned window - r

Question

So here is an example of my data
> d
customer date revenue
1: A 2016-01-01 32
2: A 2016-01-03 88
3: A 2016-01-04 80
4: A 2016-02-01 38
5: B 2016-01-13 44
6: B 2016-01-24 11
7: B 2016-01-25 50
8: B 2016-02-26 46
> dput(d)
structure(list(customer = c("A", "A", "A", "A", "B", "B", "B",
"B"), date = structure(c(16801, 16803, 16804, 16832, 16813, 16824,
16825, 16857), class = "Date"), revenue = c(32, 88, 80, 38, 44,
11, 50, 46)), .Names = c("customer", "date", "revenue"), row.names = c(NA,
-8L), class = c("data.table", "data.frame"), .internal.selfref = <pointer: 0x0000000002a60788>)
What i want to do is, I want to create a column, let call it roll_sum_3days.
This column is the rolling sum of revenue that happen afterward. The window size is conditioned on date column. In this case, roll_sum_3days is the sum of revenue that happen afterward and must not be later than 3 days.
The expected out come would look like this
customer date revenue roll_sum_3days
1: A 2016-01-01 32 168
2: A 2016-01-03 88 80
3: A 2016-01-04 80 0
4: A 2016-02-01 38 0
5: B 2016-01-13 44 0
6: B 2016-01-24 11 96
7: B 2016-01-25 50 46
8: B 2016-01-26 46 0

Answer 1

A possible solution:
library(lubridate) # for the '%m+%'-function
d[, roll_sum_3d := .SD[.SD[, .(date, date2 = date %m+% days(3), revenue)]
, on = .(date > date, date <= date2)
][, sum(revenue, na.rm = TRUE), by = date]$V1
, by = customer][]
which gives:
customer date revenue roll_sum_3d
1: A 2016-01-01 32 168
2: A 2016-01-03 88 80
3: A 2016-01-04 80 0
4: A 2016-02-01 38 0
5: B 2016-01-13 44 0
6: B 2016-01-24 11 96
7: B 2016-01-25 50 46
8: B 2016-01-26 46 0
What this does:
Group d by customer withby = customer`.
Add roll_sum_3d by reference with :=.
Calculate roll_sum_3d by joining .SD (Subset of Data) for each group with a date-window of that group (.SD[, .(date, date2 = date %m+% days(3), revenue)] with a non-equi join on = .(date > date, date <= date2), summarise the revenue for each date and give that back.
An alternative based on #Arun's comment:
d[, roll_sum_3d := d[d[, .(customer, date, date2 = date %m+% days(3), revenue)]
, on = .(customer, date > date, date <= date2)
, sum(revenue, na.rm = TRUE), by=.EACHI]$V1][]

Answer 2

Hi I guess there is another mistake in your example: observation number 8 won't add to the count of the two previous observations as it is from february. Nevermind I've got a solution if you want using apply() and the POSIXct() function
df <- data.frame(customer = c("A", "A", "A", "A", "B", "B", "B", "B"),
date = structure(c(16801, 16803, 16804, 16832, 16813, 16824,
16825, 16857), class = "Date"),
revenue = c(32, 88, 80, 38, 44, 11, 50, 46))
df$date <- as.POSIXct(df$date)
calc <- function(x){
date <- as.POSIXct(unlist(x["date"]),origin = "1970-01-01")
customer <- unlist(x["customer"])
# There you choose what you want to sum (here conditions are between the day and 3 days later and same customer)
# 86400 is the number of second in a day!
output <- sum(df[df$date > date & df$date <= (date+86400*3) & df$customer==customer,"revenue"])
return(output)
}
df$sum <- apply(df,1,calc)
# if you want to come back with your date format.
df$date <- as.Date(df$date)
df
customer date revenue sum
1 A 2016-01-01 32 168
2 A 2016-01-03 88 80
3 A 2016-01-04 80 0
4 A 2016-02-01 38 0
5 B 2016-01-13 44 0
6 B 2016-01-24 11 50
7 B 2016-01-25 50 0
8 B 2016-02-26 46 0
I couldn't keep your date format as the operator > won't work with it.