vec_order() to prepare for a breaking change in vctrs (#153).Fixed a C alignment issue detected by CRAN’s USBAN machine related to allocating vectors of long double.
Fixed a test that relied too strongly on the size of the C type, long double, which can vary across platforms (#147).
Fixed an out of sync vignette entry (#148).
New family of very fast specialized sliding functions:
slide_sum(), slide_index_sum(): for rolling sums
slide_mean(), slide_index_mean(): for rolling averages
slide_prod(), slide_index_prod(): for rolling products
slide_min(), slide_index_min(): for rolling minimums
slide_max(), slide_index_max(): for rolling maximums
slide_any(), slide_index_any(): for rolling any
slide_all(), slide_index_all(): for rolling all
The slide_index_*() family now allows .before and .after to be functions of 1 argument (the index) that compute the boundaries of the sliding window. This can be extremely useful when the default, which computes .i - .before and .i + .after, is not applicable or correct for your needs. One use case is to set .before = ~.x %m-% months(1) rather than .before = months(1) to perform a 1 month rolling window in a way that won’t generate NA values on invalid dates (like 1 month before 2019-03-31) (#139).
The slide_index_*() family has undergone some internal changes to make it more compatible with custom vctrs classes that could be provided as the index (.i), such as the date-time classes in the clock package (#133, #130).
For the slide_index_*() family, it is now required that .i - .before and .i + .after be castable to .i by vctrs::vec_cast(). Similarly, for the hop_index_*() family, .starts and .stops must both be castable to .i (#132).
New vignette, vignette("tsibble"), explaining how to transition from tsibble to slider (#128).
vignette("rowwise") has been updated to use cur_data() from dplyr 1.0.0, which makes it significantly easier to do rolling operations on data frames (like rolling regressions) using slider in a dplyr pipeline.
slide_period() and friends have slightly better handling of size zero input when .complete = TRUE (#111).
Better error messages for NA input with .before, .after, .step and .complete have been added (#110).
A few instances of possibly unsafe C protection usage have been fixed (#112).
Tests have been updated to use only numeric values in the vctrs::new_date() constructor (#113).
As a followup to a change in slider 0.1.3, edge cases with size zero input in hop() have also been fixed.
C code has been refactored to be less reliant on vctrs internals.
Updated to stay compatible with vctrs 0.3.0.
A few edge cases with size zero input in the index functions have been fixed.
The default for the .names_to argument of *_dfr() variants has been updated to rlang::zap() to match the default of the function it is passed on to, vctrs::vec_rbind().
All *_vec() variants now maintain size stability when auto-simplifying (i.e. when .ptype = NULL) (#78, #93).
hop() and its variants no longer place the names of .x on the output. Because there is no size guarantee on the output, the size of .x can be different than the size of the output, meaning that the names might not line up. This also affects slide_period(), which is implemented using a hop() variant (#75).
With data frames containing row names, slide() and its variants now copy those row names onto the output. This is an implicit benefit from vctrs gaining better support for data frame row names.
NEWS.md file to track changes to the package.