`runner`

package provides functions applied on running
windows. The most universal function is `runner::runner`

which gives user possibility to apply any R function `f`

on
running windows. Running windows are defined for each data window size
`k`

, `lag`

with respect to their indexes. Unlike
other available R packages, `runner`

supports any input and
output type and also gives full control to manipulate window size and
lag/lead.

There are different kinds of running windows and all of them are
implemented in `runner`

.

The simplest window type which is similar to
`base::cumsum`

. At each element window is defined by all
elements appearing before current.

In `runner`

this can be achieved as simple by:

```
library(runner)
# full windows
runner(1:15)
# summarizing - sum
runner(
1:15,
f = sum
)
# summarizing - concatenating
runner(
1:15],
letters[f = paste,
collapse = " > "
)
```

Second type of windows are these commonly known as
running/rolling/moving/sliding windows. This types of windows moves
along the index instead of cumulating like a previous one.

Following diagram illustrates running windows of length
`k = 4`

. Each of 15 windows contains 4 elements (except first
three).

To obtain constant sliding windows one just needs to specify
`k`

argument

```
# summarizing - sum of 4-elements
runner(
1:15,
k = 4,
f = sum
)
# summarizing - slope from lm
<- data.frame(
df a = 1:15,
b = 3 * 1:15 + rnorm(15)
)
runner(
x = df,
k = 5,
f = function(x) {
<- lm(b ~ a, data = x)
model coefficients(model)["a"]
} )
```

By default `runner`

calculates on assumption that index
increments by one, but sometimes data points in dataset are not equally
spaced (missing weekends, holidays, other missings) and thus window size
should vary to keep expected time frame. If one specifies
`idx`

argument, than running functions are applied on windows
depending on date rather on a sequence 1-n. `idx`

should be
the same length as `x`

and should be of type
`Date`

, `POSIXt`

or `integer`

. Example
below illustrates window of size `k = 5`

lagged by
`lag = 1`

. Note that one can specify also
`k = "5 days"`

and `lag = "day"`

as in
`seq.POSIXt`

.

In the example below in square brackets ranges for each window.

```
<- c(4, 6, 7, 13, 17, 18, 18, 21, 27, 31, 37, 42, 44, 47, 48)
idx
# summarize - mean
::runner(
runnerx = idx,
k = 5, # 5-days window
lag = 1,
idx = idx,
f = function(x) mean(x)
)
# use Date or datetime sequences
::runner(
runnerx = idx,
k = "5 days", # 5-days window
lag = 1,
idx = Sys.Date() + idx,
f = function(x) mean(x)
)
# obtain window from above illustration
::runner(
runnerx = idx,
k = "5 days",
lag = 1,
idx = Sys.Date() + idx
)
```

Runner by default returns vector of the same size as `x`

unless one puts any-size vector to `at`

argument. Each
element of `at`

is an index on which runner calculates
function. Example below illustrates output of runner for
`at = c(13, 27, 45, 31)`

which gives windows in ranges
enclosed in square brackets. Range for `at = 27`

is
`[22, 26]`

which is not available in current indices.

```
<- c(4, 6, 7, 13, 17, 18, 18, 21, 27, 31, 37, 42, 44, 47, 48)
idx
# summary
::runner(
runnerx = 1:15,
k = 5,
lag = 1,
idx = idx,
at = c(18, 27, 48, 31),
f = mean
)
# full window
::runner(
runnerx = idx,
k = 5,
lag = 1,
idx = idx,
at = c(18, 27, 48, 31)
)
```

`at`

can also be specified as interval of the output
defined by time interval which results in obtaining results on following
indices
`seq(min(idx), max(idx), by = "<time interval>")`

.
Interval can be set in the same way as in `seq.POSIXt`

function. It’s worth noting that `at`

interval shouldn’t be
more frequent than interval of `idx`

- for `Date`

the most frequent interval is a `"day"`

, for
`POSIXt`

it’s a `"sec"`

.

```
<- seq(Sys.Date(), Sys.Date() + 365, by = "1 month")
idx_date
# change interval to 4-months
runner(
x = 0:12,
idx = idx_date,
at = "4 months"
)
# calculate correlation at every 6-months
runner(
x = data.frame(
a = 1:13,
b = 1:13 + rnorm(13, sd = 5),
idx_date
),idx = "idx_date",
at = "6 months",
f = function(x) {
cor(x$a, x$b)
} )
```

One can stretch window length by `k`

and shift in time (or
index) using `lag`

. Both arguments can be
`integer`

and also time interval like for example
`2 months`

. If `k`

or `lag`

are a
single value then window size/lag are constant for all elements of x.
User can also specify `k/lag`

as vector, then size and lag
will vary for each window. Both `k`

and `lag`

can
be of `length(.) == 1`

, `length(.) == length(x)`

or `length(.) == length(at)`

(if `at`

is
specified). `lag`

can be negative and positive while
`k`

only non-negative.

```
# summarizing - concatenating
::runner(
runnerx = 1:10,
lag = c(-1, 2, -1, -2, 0, 0, 5, -5, -2, -3),
k = c(0, 1, 1, 1, 1, 5, 5, 5, 5, 5),
f = paste,
collapse = ","
)
# full window
::runner(
runnerx = 1:10,
lag = 1,
k = c(1, 1, 1, 1, 1, 5, 5, 5, 5, 5)
)
# on dates
<- c(4, 6, 7, 13, 17, 18, 18, 21, 27, 31, 37, 42, 44, 47, 48)
idx
::runner(
runnerx = 1:15,
lag = sample(c("-2 days", "-1 days", "1 days", "2 days"),
size = 15,
replace = TRUE
),k = sample(c("5 days", "10 days", "15 days"),
size = 15,
replace = TRUE
),idx = Sys.Date() + idx,
f = function(x) mean(x)
)
```

`NA`

paddingUsing `runner`

one can also specify
`na_pad = TRUE`

which would return `NA`

for any
window which is partially out of range - meaning that there is no
sufficient number of observations to fill the window. By default
`na_pad = FALSE`

, which means that incomplete windows are
calculated anyway. `na_pad`

is applied on normal cumulative
windows and on windows depending on date. In example below two windows
exceed range given by `idx`

so for these windows are empty
for `na_pad = TRUE`

. If used sets `na_pad = FALSE`

first window will be empty (no single element within
`[-2, 3]`

) and last window will return elements within
matching `idx`

.

```
<- c(4, 6, 7, 13, 17, 18, 18, 21, 27, 31, 37, 42, 44, 47, 48)
idx
::runner(
runnerx = 1:15,
k = 5,
lag = 1,
idx = idx,
at = c(4, 18, 48, 51),
na_pad = TRUE,
f = function(x) mean(x)
)
```

`data.frame`

User can also put `data.frame`

into `x`

argument and apply functions which involve multiple columns. In example
below we calculate beta parameter of `lm`

model on 1, 2, …, n
observations respectively. On the plot one can observe how
`lm`

parameter adapt with increasing number of
observation.

```
<- cumsum(rnorm(40))
x <- 3 * x + rnorm(40)
y <- Sys.Date() + cumsum(sample(1:3, 40, replace = TRUE)) # unequaly spaced time series
date <- rep(c("a", "b"), 20)
group
<- data.frame(date, group, y, x)
df
<- runner(
slope
df,function(x) {
coefficients(lm(y ~ x, data = x))[2]
}
)
plot(slope)
```

One can also use `runner`

with `dplyr`

also
with problematic `group_by`

operations, without need to apply
group_modify.
Below we apply grouped 20-days beta, by specifying window length
`k = "10 days"`

and providing column name where indices
(dates) are kept.

```
library(dplyr)
<- df %>%
summ group_by(group) %>%
mutate(
cumulative_mse = runner(
x = .,
k = "20 days",
idx = "date", # specify column name instead df$date
f = function(x) {
coefficients(lm(y ~ x, data = x))[2]
}
)
)
library(ggplot2)
%>%
summ ggplot(aes(x = date, y = cumulative_mse, group = group, color = group)) +
geom_line()
```

When user executes multiple `runner`

calls in
`dplyr`

mutate, one can also use `run_by`

function
to prespecify arguments in `tidyverse`

pipeline. In the
example below `runner`

functions are applied on
`k = "20 days"`

calculated on `"date"`

column.

```
%>%
df group_by(group) %>%
run_by(idx = "date", k = "20 days", na_pad = FALSE) %>%
mutate(
cumulative_mse = runner(
x = .,
f = function(x) {
mean((residuals(lm(y ~ x, data = x)))^2)
}
),intercept = runner(
x = .,
f = function(x) {
coefficients(lm(y ~ x, data = x))[1]
}
),slope = runner(
x = .,
f = function(x) {
coefficients(lm(y ~ x, data = x))[2]
}
) )
```

The `runner`

function can also compute windows in parallel
mode. The function doesn’t initialize the parallel cluster automatically
but one have to do this outside and pass it to the `runner`

through `cl`

argument.

```
library(parallel)
<- detectCores()
numCores <- makeForkCluster(numCores)
cl
runner(
x = df,
k = 10,
idx = "date",
f = function(x) sum(x$x),
cl = cl
)
stopCluster(cl)
```

*Executing runner in parallel mode isn’t always
faster than a single thread.*

With `runner`

one can use any R functions, but some of
them are optimized for speed reasons. These functions are:

- aggregating functions - `length_run`

, `min_run`

,
`max_run`

, `minmax_run`

, `sum_run`

,
`mean_run`

, `streak_run`

- utility functions - `fill_run`

, `lag_run`

,
`which_run`