Why sum and mean works different in apply.monthly?

[fcasarramona]

When apply.monthly in a xts multivaluate time series (with more than one column) apply.monthly(xts.ts, sum) returns a one column time series but apply.monthly(xts.ts, mean) returns the same number of columns of the original.

Here is an example:

library(xts)
xts.ts <- xts(matrix(rnorm(462), nrow = 231),as.Date(13514:13744,origin="1970-01-01"))
dim(xts.ts)
[1] 231   2
dim(apply.monthly(xts.ts, mean))
[1] 8 2
dim(apply.monthly(xts.ts, sum))
[1] 8 1

The same is true with other xts apply functions.

This seems inconsistent.

[braverock]

Why not use colMeans and colSums to make sure you get what you think you're getting?

x<-xts(matrix(rnorm(1000),ncol=10),order.by=as.Date(1:100))
mean(x)
sum(x)
apply.weekly(x,sum)
apply.weekly(x,mean)
apply.weekly(x,colSums)
apply.weekly(x,colMeans)

[fcasarramona]

Thanks, just discovered I can do it with colSums.
In this case, apply.weekly(x,sum) and apply.weekly(x,mean) should return 1 column xts both.

[joshuaulrich]

The problem is that xts defines mean.xts in its namespace and it is being dispatched, even though it is not registered as a S3 method.

mean.xts was added when R Core decided that mean should not operate by-column for matrices and data.frames. Seems like it might be a good idea to rename both mean.xts and sd.xts as mean_xts and sd_xts, respectively, in order to avoid this inconsistency.

Or maybe they should just be removed all together? They're not exported, so CRAN likely would not allow their use in other packages.

[fcasarramona]

In my opinion is better just remove them and let colSums and his family functions do the work. This will be consistent with matrix.

[sushant-choudhary]

Using colSums takes care of the sum function not being applied by columns, however what would be a similar fix for the sd( ) function, which, like sum( ), is not applied by columns in period.apply.

[pverspeelt]

@sushant-choudhary ,

A late answer, but the package matrixStats has all these functions available for columns and rows. Using the example above that would be:

apply.weekly(x, matrixStats::colSds)

[joshuaulrich]

I just pushed a commit with the best solution I could think of. The current behavior is maintained, but prints a message every time period.apply(x, ep, FUN = mean) is called, unless options(xts.message.period.apply.mean = FALSE). This is a message instead of a warning because a warning would cause CRAN packages to fail R CMD check.

The message says to use FUN = colMeans for the current behavior and FUN = function(x) mean(x) to calculate the mean for all the data. Neither of those will print a message. I also added a note about this to the help files for period.apply() and apply.daily() (and friends).