slide_index: Slide relative to an index

View source: R/slide-index.R

slide_indexR Documentation

Slide relative to an index


slide_index() is similar to slide(), but allows a secondary .i-ndex vector to be provided.

This is often useful in business calculations, when you want to compute a rolling computation looking "3 months back", which is approximately but not equivalent to, 3 * 30 days. slide_index() allows for these irregular window sizes.


slide_index(.x, .i, .f, ..., .before = 0L, .after = 0L, .complete = FALSE)

  .before = 0L,
  .after = 0L,
  .complete = FALSE,
  .ptype = NULL

slide_index_dbl(.x, .i, .f, ..., .before = 0L, .after = 0L, .complete = FALSE)

slide_index_int(.x, .i, .f, ..., .before = 0L, .after = 0L, .complete = FALSE)

slide_index_lgl(.x, .i, .f, ..., .before = 0L, .after = 0L, .complete = FALSE)

slide_index_chr(.x, .i, .f, ..., .before = 0L, .after = 0L, .complete = FALSE)

  .before = 0L,
  .after = 0L,
  .complete = FALSE,
  .names_to = rlang::zap(),
  .name_repair = c("unique", "universal", "check_unique")

  .before = 0L,
  .after = 0L,
  .complete = FALSE,
  .size = NULL,
  .name_repair = c("unique", "universal", "check_unique", "minimal")




The vector to iterate over and apply .f to.



The index vector that determines the window sizes. It is fairly common to supply a date vector as the index, but not required.

There are 3 restrictions on the index:

  • The size of the index must match the size of .x, they will not be recycled to their common size.

  • The index must be an increasing vector, but duplicate values are allowed.

  • The index cannot have missing values.


⁠[function / formula]⁠

If a function, it is used as is.

If a formula, e.g. ~ .x + 2, it is converted to a function. There are three ways to refer to the arguments:

  • For a single argument function, use .

  • For a two argument function, use .x and .y

  • For more arguments, use ..1, ..2, ..3 etc

This syntax allows you to create very compact anonymous functions.


Additional arguments passed on to the mapped function.

.before, .after

⁠[vector(1) / function / Inf]⁠

  • If a vector of size 1, these represent the number of values before or after the current element of .i to include in the sliding window. Negative values are allowed, which allows you to "look forward" from the current element if used as the .before value, or "look backwards" if used as .after. Boundaries are computed from these elements as .i - .before and .i + .after. Any object that can be added or subtracted from .i with + and - can be used. For example, a lubridate period, such as lubridate::weeks().

  • If Inf, this selects all elements before or after the current element.

  • If a function, or a one-sided formula which can be coerced to a function, it is applied to .i to compute the boundaries. Note that this function will only be applied to the unique values of .i, so it should not rely on the original length of .i in any way. This is useful for applying a complex arithmetic operation that can't be expressed with a single - or + operation. One example would be to use lubridate::add_with_rollback() to avoid invalid dates at the end of the month.

The ranges that result from applying .before and .after have the same 3 restrictions as .i itself.



Should the function be evaluated on complete windows only? If FALSE, the default, then partial computations will be allowed.


⁠[vector(0) / NULL]⁠

A prototype corresponding to the type of the output.

If NULL, the default, the output type is determined by computing the common type across the results of the calls to .f.

If supplied, the result of each call to .f will be cast to that type, and the final output will have that type.

If getOption("vctrs.no_guessing") is TRUE, the .ptype must be supplied. This is a way to make production code demand fixed types.


This controls what to do with input names supplied in ....

  • By default, input names are zapped.

  • If a string, specifies a column where the input names will be copied. These names are often useful to identify rows with their original input. If a column name is supplied and ... is not named, an integer column is used instead.

  • If NULL, the input names are used as row names.


One of "unique", "universal", "check_unique", "unique_quiet", or "universal_quiet". See vec_as_names() for the meaning of these options.

With vec_rbind(), the repair function is applied to all inputs separately. This is because vec_rbind() needs to align their columns before binding the rows, and thus needs all inputs to have unique names. On the other hand, vec_cbind() applies the repair function after all inputs have been concatenated together in a final data frame. Hence vec_cbind() allows the more permissive minimal names repair.


If, NULL, the default, will determine the number of rows in vec_cbind() output by using the tidyverse recycling rules.

Alternatively, specify the desired number of rows, and any inputs of length 1 will be recycled appropriately.


A vector fulfilling the following invariants:


  • vec_size(slide_index(.x)) == vec_size(.x)

  • vec_ptype(slide_index(.x)) == list()

slide_index_vec() and ⁠slide_index_*()⁠ variants

  • vec_size(slide_index_vec(.x)) == vec_size(.x)

  • vec_size(slide_index_vec(.x)[[1]]) == 1L

  • vec_ptype(slide_index_vec(.x, .ptype = ptype)) == ptype

See Also

slide(), hop_index(), slide_index2()



x <- 1:5

# In some cases, sliding over `x` with a strict window size of 2
# will fit your use case.
slide(x, ~.x, .before = 1)

# However, if this `i` is a date vector paired with `x`, when computing
# rolling calculations you might want to iterate over `x` while
# respecting the fact that `i` is an irregular sequence.
i <- as.Date("2019-08-15") + c(0:1, 4, 6, 7)

# For example, a "2 day" window should not pair `"2019-08-19"` and
# `"2019-08-21"` together, even though they are next to each other in `x`.
# `slide_index()` computes the lookback value from the current date in `.i`,
# meaning that if you are currently on `"2019-08-21"` and look back 1 day,
# it will correctly not include `"2019-08-19"`.
slide_index(i, i, ~.x, .before = 1)

# We could have equivalently used a lubridate period object for this as well,
# since `i - lubridate::days(1)` is allowed
slide_index(i, i, ~.x, .before = lubridate::days(1))

# ---------------------------------------------------------------------------
# Functions for `.before` and `.after`

# In some cases, it might not be appropriate to compute
# `.i - .before` or `.i + .after`, either because there isn't a `-` or `+`
# method defined, or because there is an alternative way to perform the
# arithmetic. For example, subtracting 1 month with `- months(1)` (using
# lubridate) can sometimes land you on an invalid date that doesn't exist.
i <- as.Date(c("2019-01-31", "2019-02-28", "2019-03-31"))

# 2019-03-31 - months(1) = 2019-02-31, which doesn't exist
i - months(1)

# These NAs create problems with `slide_index()`, which doesn't allow
# missing values in the computed endpoints
try(slide_index(i, i, identity, .before = months(1)))

# In these cases, it is more appropriate to use `%m-%`,
# which will snap to the end of the month, at least giving you something
# to work with.
i %m-% months(1)

# To use this as your `.before` or `.after`, supply an anonymous function of
# 1 argument that performs the computation
slide_index(i, i, identity, .before = ~.x %m-% months(1))

# Notice that in the `.after` case, `2019-02-28 %m+% months(1)` doesn't
# capture the end of March, so it isn't included in the 2nd result
slide_index(i, i, identity, .after = ~.x %m+% months(1))

# ---------------------------------------------------------------------------

# When `.i` has repeated values, they are always grouped together.
i <- c(2017, 2017, 2018, 2019, 2020, 2020)
slide_index(i, i, ~.x)
slide_index(i, i, ~.x, .after = 1)

# ---------------------------------------------------------------------------
# Rolling regressions

# Rolling regressions are easy with `slide_index()` because:
# - Data frame `.x` values are iterated over rowwise
# - The index is respected by using `.i`

df <- data.frame(
  y = rnorm(100),
  x = rnorm(100),
  i = as.Date("2019-08-15") + c(0, 2, 4, 6:102) # <- irregular

# 20 day rolling regression. Current day + 19 days back.
# Additionally, set `.complete = TRUE` to not compute partial results.
regr <- slide_index(df, df$i, ~lm(y ~ x, .x), .before = 19, .complete = TRUE)


# The first 16 slots are NULL because there is no possible way to
# look back 19 days from the 16th index position and construct a full
# window. But on the 17th index position, `""2019-09-03"`, if we look
# back 19 days we get to `""2019-08-15"`, which is the same value as
# `i[1]` so a full window can be constructed.
df$i[16] - 19 >= df$i[1] # FALSE
df$i[17] - 19 >= df$i[1] # TRUE

# ---------------------------------------------------------------------------
# Accessing the current index value

# A very simplistic version of `purrr::map2()`
fake_map2 <- function(.x, .y, .f, ...) {
  Map(.f, .x, .y, ...)

# Occasionally you need to access the index value that you are currently on.
# This is generally not possible with a single call to `slide_index()`, but
# can be easily accomplished by following up a `slide_index()` call with a
# `purrr::map2()`. In this example, we want to use the distance from the
# current index value (in days) as a multiplier on `x`. Values further
# away from the current date get a higher multiplier.

# 25 random days past 2000-01-01
i <- sort(as.Date("2000-01-01") + sample(100, 25))

df <- data.frame(i = i, x = rnorm(25))

weight_by_distance <- function(df, i) {
  df$weight = abs(as.integer(df$i - i))
  df$x_weighted = df$x * df$weight

# Use `slide_index()` to just generate the rolling data.
# Here we take the current date + 5 days before + 5 days after.
dfs <- slide_index(df, df$i, ~.x, .before = 5, .after = 5)

# Follow up with a `map2()` with `i` as the second input.
# This allows you to track the current `i` value and weight accordingly.
result <- fake_map2(dfs, df$i, weight_by_distance)


slider documentation built on Oct. 12, 2023, 5:11 p.m.