recipes step_window() in R: Smooth Numeric Predictors

The recipes step_window() function in R computes a moving-window statistic, such as a rolling mean or median, across a numeric column to smooth out short-term noise. You add it to a recipe(), register it with prep(), and apply it with bake().

step_window(rec, units)                        # 3-point moving mean, in place
step_window(rec, units, size = 5)              # wider 5-point window
step_window(rec, units, statistic = "median")  # rolling median, robust to spikes
step_window(rec, units, statistic = "sum")     # rolling sum over the window
step_window(rec, units, names = "units_ma")    # new column, keep the original
step_window(rec, units, na_rm = FALSE)         # let NA propagate into windows
prep(rec) |> bake(new_data = NULL)             # learn, then apply

What step_window() does in R

step_window() replaces a column with its own moving-window summary. Raw measurements often jitter from week to week even when the underlying trend is smooth. step_window() slides a fixed-width window along a numeric column and, at every row, replaces the value with a statistic computed from the values inside that window. The default statistic is the mean, so the column becomes a moving average of itself.

It belongs to the recipes package, the feature-engineering layer of tidymodels. The window is center-justified: for a window of size = 3, the smoothed value at a row is the statistic of that row plus its immediate neighbour on each side. The first and last rows have no full window, so they reuse the first and last complete window values respectively. That keeps the output the same length as the input with no NA cells.

step_window() syntax and arguments

step_window() attaches a windowing operation to a recipe. You pass the recipe first, then the numeric columns to smooth, named directly or chosen with a selector such as all_numeric_predictors().

The arguments you will actually touch:

The size and statistic arguments shape the smoothing. A larger size averages over more rows and produces a flatter line, while statistic swaps the mean for a median, sum, or spread measure. Both are common targets for tuning when the smoothing window itself is a model hyperparameter.

step_window() examples

Build the recipe, prep it, then bake. A recipe is only a plan until prep() registers the steps and bake() applies them. This recipe smooths units with the default 3-point moving mean.

The zigzag in the raw units figures is gone, leaving a clean rising trend. Notice the column is still called units: the smoothed values replaced the originals. To keep both, pass the names argument.

Now units keeps its raw values and units_ma carries the moving average alongside it. The size and statistic arguments change how aggressively the column is smoothed.

A size = 5 window spans five rows, and statistic = "median" reports the middle value instead of the average. The median ignores how far an outlier sits from the centre, so a single spike moves it far less than it moves a mean.

step_window() vs step_lag() vs slider

Pick the tool that matches whether you want an aggregate, a copy, or raw control. A window summarises nearby rows, a lag copies one past row, and a manual rolling call trades convenience for flexibility.

Use step_window() when a smoothed trend matters more than any single observation. Reach for step_lag() when the exact past value is the feature you need. Drop down to slider only for a rolling computation that statistic cannot express.

Common pitfalls with step_window()

Most failures trace back to window size or row order. Watch these three traps when adding the step.

1. Even or too-small window. size must be an odd integer of 3 or more. Passing size = 4 or size = 2 throws an error before the recipe preps, because a center-justified window has no defined midpoint without an odd width.
2. Unsorted data. The window slides by row position, not by date. If the rows are not already in time order, the moving average mixes unrelated periods. Sort with arrange(week) before you build the recipe.
3. Mixing several series in one frame. step_window() has no group awareness. With stacked panels, such as sales for many stores, the window straddles the boundary between stores. Smooth each series separately or filter to one before applying the step.

Try it yourself

Related recipes steps

step_window() is one of several recipes steps for time-based predictors. These pair naturally with it in a tidymodels workflow:

• step_lag() creates shifted copies of a column as past-value features.
• step_date() extracts month, day-of-week, and year from a date.
• step_impute_roll() fills missing values with a rolling-window statistic.
• step_normalize() centers and scales a numeric column.
• recipe() is the starting point every step is added to.

See the official step_window() reference for the full argument list.

FAQ

What does step_window() do in R?

step_window() is a recipes step that computes a moving-window statistic across a numeric column. At every row it slides a fixed-width window, summarises the values inside with a statistic such as the mean or median, and uses that as the row's value. The default is a 3-point moving mean. You add it to a recipe(), register it with prep(), and apply it with bake(). By default it replaces the column in place, keeping the output the same length as the input.

What window sizes can step_window() use?

The size argument must be an odd integer of 3 or more. Odd widths are required because the window is center-justified: each smoothed value sits at the midpoint of its window, and only an odd count of rows has a true middle. Common choices are 3, 5, and 7. A larger size produces a flatter, less responsive line and is often tuned as a hyperparameter.

Does step_window() replace the original column?

Yes, by default. With no names argument, the smoothed values overwrite the source column and the role stays unchanged. To keep the raw column, pass names with one new name per selected column. The step then writes those new columns and leaves the originals in place, controlled by keep_original_cols, which defaults to TRUE.

Why does step_window() need RcppRoll?

The rolling calculations are delegated to the RcppRoll package, which implements them in compiled C++ for speed. recipes lists RcppRoll as a suggested dependency rather than a hard one, so it is not installed automatically. If you call step_window() without it, prep() stops with a message telling you to run install.packages("RcppRoll"). Install it once and the step works normally.