slide provides a family of general purpose “sliding window” functions. The API is purposefully very similar to purrr. The goal of these functions is usually to compute rolling averages, cumulative sums, rolling regressions, or other “window” based computations.

There are 3 core functions in slide:

  • slide() iterates over your data like purrr::map(), but uses a sliding window to do so. It is type stable, and always returns a result with the same size as its input.

  • slide_index() computes a rolling calculation relative to an index. If you have ever wanted to compute something like a “3 month rolling average” where the number of days in each month is irregular, you might like this function.

  • slide_between() is a lower level version of slide_index() that allows you to manually specify sliding boundaries using .starts and .stops. It can be useful if you want to return a result with a different size from the original input, or want to aggregate data to a less granular frequency using a rolling window (like taking daily data, slicing it into monthly blocks, and then creating a rolling window over those monthly blocks).

Each of these 3 functions has the same variants as purrr::map(). For example, slide() has slide_dbl(), slide2(), and pslide(), along with the other combinations of these variants that you might expect from having previously used purrr.

To learn more about these three functions, read the introduction vignette.

Installation

slide is NOT yet on CRAN.

You can install the development version from GitHub with:

remotes::install_github("DavisVaughan/slide")

Examples

The help page for slide() has many examples, but here are a few:

library(slide)

The classic example would be to do a moving average. slide() handles this with a combination of the .before and .after arguments, which control the width of the window and the alignment.

With Inf, you can do a “cumulative slide” to compute cumulative expressions. I think of this as saying “give me everything before the current element.”

With .complete, you can decide whether or not .f should be evaluated on incomplete windows. In the following example, the requested window size is 3, but the first two results are computed on windows of size 1 and 2 because partial results are allowed by default. When .complete is set to TRUE, the first two results are not computed.

Index sliding

In many business settings, the value you want to compute is tied to some index, like a date vector. In these cases, you’ll probably want to compute sliding windows relative to the index, and not using the fixed window that slide() provides. You can use slide_index() to pass in both .x and an index, .i, and the window will be calculated relative to that index.

Here, when computing a “2 day window”, you probably don’t want "2019-08-16" and "2019-08-18" to be grouped together. slide() has no concept of an index, so when you specify a window size of 2, it will group these two together. slide_index(), on the other hand, will do the right thing.

Essentially what happens is that when we get to "2019-08-18", it “looks backwards” 1 day to set a window boundary at "2019-08-17". Since the date at position 2, "2019-08-16", is before "2019-08-17", it is not included.

Powerfully, you can pass through any object to .before that computes a value from .i - .before. This means that you could also have used a lubridate period object (which gets even more interesting when you use weeks() or months()):

Inspiration

This package is inspired heavily by SQL’s window functions. The API is similar, but more general because you can iterate over any kind of R object.

There have been multiple attempts at creating sliding window functions (I personally created rollify(), and worked a little bit on tsibble::slide() with Earo Wang).

I believe that slide is the next iteration of these. There are a few reasons for this:

  • To me, the API is more intuitive, and is more flexible because .before and .after let you completely control the entry point (as opposed to fixed entry points like "center", "left", etc.

  • It is objectively faster because it is written purely in C.

  • With slide_vec() you can return any kind of object, and are not limited to the suffixed versions: _dbl, _int, etc.

  • It iterates rowwise over data frames, consistent with the vctrs framework.

  • I believe it is overall more consistent, backed by a theory that can always justify the sliding window generated by any combination of the parameters.

Earo and I have spoken, and we have mututally agreed that it would be best to deprecate tsibble::slide() in favor of slide::slide().

Additionally, data.table’s non-equi joins have been pretty much the only solution to the problem that slide_index() tries to solve. Their solution is robust and quite fast, and has been a nice benchmark for slide. slide is trying to solve a much narrower problem, so the API here is more focused.

Performance

In terms of performance, be aware that any specialized package that shifts the function calls to C are going to be faster than slide. For example, RcppRoll::roll_mean() computes the rolling mean at the C level, which is bound to be faster. The purpose of slide is to be general purpose, while still being as fast as possible. This means that it can be used for more abstract things, like rolling regressions, or any other custom function that you want to use in a rolling fashion.

Otherwise, like purrr::map(), slide() is optimized in C to be as fast as possible, getting out of the way as quickly as it can so the main overhead are the .f calls.

References

I’ve found the following references very useful to understand more about window functions: