Writing ts-functions

It is straightforward to turn existing functions into functions that can deal with any ts-boxable object.

The ts_ function is a constructor function for tsbox time series functions. It can be used to wrap any function that works with time series. The default is set to R base "ts" class, so wrapping functions for "ts" time series (or vectors or matrices) is as simple as:

ts_rowsums <- ts_(rowSums)
ts_rowsums(ts_c(mdeaths, fdeaths))

Note that ts_ returns a function, which can be with or without a name. Let’ have a closer look at how ts_rowsums looks like:

ts_rowsums
#> function (x, ...) 
#> {
#>     stopifnot(ts_boxable(x))
#>     z <- rowSums(ts_ts(x), ...)
#>     copy_class(z, x)
#> }

This is how most ts-functions work. They use a specific converter function (here: ts_ts) to convert a ts-boxable object to the desired class. They then perform the main operation on the object (here: rowSums). Finally they convert the result back to the original class, using copy_class.

Functions from external packages

Here is a slightly more complex example, which uses a post processing function:

ts_prcomp <- ts_(function(x) predict(prcomp(x, scale = TRUE)))
ts_prcomp(ts_c(mdeaths, fdeaths))

It is easy to make functions from external packages ts-boxable, by wrapping them into ts_.

ts_dygraphs <- ts_(dygraphs::dygraph, class = "xts")
ts_forecast <- ts_(function(x) forecast::forecast(x)$mean, vectorize = TRUE)
ts_seas <- ts_(function(x) seasonal::final(seasonal::seas(x)), vectorize = TRUE)

ts_dygraphs(ts_c(mdeaths, EuStockMarkets))
ts_forecast(ts_c(mdeaths, fdeaths))
ts_seas(ts_c(mdeaths, fdeaths))

If you are explicit about the namespace (e.g., dygraphs::dygraph), ts_ recognized the package in use and delivers a meaningful message if the package is not installed.

Note that the ts_ function deals with the conversion stuff, ‘vectorizes’ the function so that it can be used with multiple time series.

Let’ have another look at ts_forecast:

ts_forecast
#> function (x, ...) 
#> {
#>     load_suggested("forecast")
#>     ff <- function(x, ...) {
#>         stopifnot(ts_boxable(x))
#>         z <- (function(x) forecast::forecast(x)$mean)(ts_ts(x), 
#>             ...)
#>         copy_class(z, x)
#>     }
#>     ts_apply(x, ff, ...)
#> }

There three main differences here: First, the function requires the forecast package. If it is not installed, load_suggested will ask the user to do so. Second, the function in use is an anonymous function, function(x) forecast::forecast(x)$mean, that also extracts the mean from the result. Third, the function is ‘vectorized’, using ts_apply. This causes the process to be repeated for each time series in the object.