What is an efficient method for partitioning and aggregating intervals from timestamped rows in a data frame?

Question

From a data frame with timestamped rows (strptime results), what is the best method for aggregating statistics for intervals?

Intervals could be an hour, a day, etc.

There's the aggregate function, but that doesn't help with assigning each row to an interval. I'm planning on adding a column to the data frame that denotes interval and using that with aggregate, but if there's a better solution it'd be great to hear it.

Thanks for any pointers!

---

Example Data

Five rows with timestamps divided into 15-minute intervals starting at 03:00.

Interval 1

• "2010-01-13 03:02:38 UTC"
• "2010-01-13 03:08:14 UTC"
• "2010-01-13 03:14:52 UTC"

Interval 2

• "2010-01-13 03:20:42 UTC"
• "2010-01-13 03:22:19 UTC"

---

Conclusion

Using a time series package such as xts should be the solution; however I had no success using them and winded up using cut. As I presently only need to plot histograms, with rows grouped by interval, this was enough.

cut is used liked so:

interv <- function(x, start, period, num.intervals) {
  return(cut(x, as.POSIXlt(start)+0:num.intervals*period))
}

Answer 1

Standard functions to split vectors are cut and findInterval:

v <- as.POSIXct(c(
  "2010-01-13 03:02:38 UTC",
  "2010-01-13 03:08:14 UTC",
  "2010-01-13 03:14:52 UTC",
  "2010-01-13 03:20:42 UTC",
  "2010-01-13 03:22:19 UTC"
))

# Your function return list:
interv(v, as.POSIXlt("2010-01-13 03:00:00 UTC"), 900)
# [[1]]
# [1] "2010-01-13 03:00:00"
# [[2]]
# [1] "2010-01-13 03:00:00"
# [[3]]
# [1] "2010-01-13 03:00:00"
# [[4]]
# [1] "2010-01-13 03:15:00 CET"
# [[5]]
# [1] "2010-01-13 03:15:00 CET"

# cut returns factor, you must provide proper breaks:
cut(v, as.POSIXlt("2010-01-13 03:00:00 UTC")+0:2*900)
# [1] 2010-01-13 03:00:00 2010-01-13 03:00:00 2010-01-13 03:00:00
# [4] 2010-01-13 03:15:00 2010-01-13 03:15:00
# Levels: 2010-01-13 03:00:00 2010-01-13 03:15:00

# findInterval returns vector of interval id (breaks like in cut)
findInterval(v, as.POSIXlt("2010-01-13 03:00:00 UTC")+0:2*900)
# [1] 1 1 1 2 2

For the record: cut has a method for POSIXt type, but unfortunately there is no way to provide start argument, effect is:

cut(v,"15 min")
# [1] 2010-01-13 03:02:00 2010-01-13 03:02:00 2010-01-13 03:02:00
# [4] 2010-01-13 03:17:00 2010-01-13 03:17:00
# Levels: 2010-01-13 03:02:00 2010-01-13 03:17:00

As you see it's start at 03:02:00. You could mess with labels of output factor (convert labels to time, round somehow and convert back to character).

Answer 2

Use a time series package. The xts package has functions designed specifically to do that. Or look at the aggregate and rollapply functions in the zoo package.

The rmetrics ebook has a useful discussion, including a performance comparison of the various packages: https://www.rmetrics.org/files/freepdf/TimeSeriesFAQ.pdf

Edit: Look at my answer to this question. Basically you need to truncate every timestamp into a specific interval and then do the aggregation using those new truncated timestamps as your grouping vector.

Answer 3

This is an interesting question; with the proliferation of the various time series packages and methods, there ought to be an approach for binning irregular time series other than by brute force that the OP suggests. Here is one "high-level" way to get the intervals that you can then use for aggregate et al, using a version of cut defined for chron objects.

require(chron)
require(timeSeries)

my.times <- "
2010-01-13 03:02:38 UTC
2010-01-13 03:08:14 UTC
2010-01-13 03:14:52 UTC
2010-01-13 03:20:42 UTC
2010-01-13 03:22:19 UTC
"

time.df <- read.delim(textConnection(my.times),header=FALSE,sep="\n",strip.white=FALSE)
time.seq <- seq(trunc(timeDate(time.df[1,1]),units="hours"),by=15*60,length=nrow(time.df))
intervals <- as.numeric(cut(as.chron(as.character(time.df$V1)),breaks=as.chron(as.character(time.seq))))

You get

intervals  
[1] 1 1 1 2 2

which you can now append to the data frame and aggregate.

The coersion acrobatics above (from character to timeDate to character to chron) is a little unfortunate, so if there are cleaner solutions for binning irregular time data using xts or any of the other timeSeries packages, I'd love to hear about them as well!..

I am also curious to know what would be the most efficient approach for binning large high-frequency irregular time series, e.g. creating 1-minute volume bars on tick data for a very liquid stock.