rsample sliding_period() in R: Calendar-Based Resamples

The rsample sliding_period() function in R builds time-series resamples whose analysis and assessment windows align to calendar units (years, months, weeks, days), so every split contains a whole number of months or weeks instead of a fixed row count and irregular gaps in the data never distort the window size.

sliding_period(df, date, "month")                                # 1-month train, 1-month test
sliding_period(df, date, "month", lookback = 11)                 # 12-month rolling train
sliding_period(df, date, "month", lookback = Inf)                # expanding train window
sliding_period(df, date, "month", assess_stop = 3)               # 3-month forecast horizon
sliding_period(df, date, "week", every = 2)                      # group rows into 2-week blocks
sliding_period(df, date, "month", step = 3)                      # advance origin 3 months
sliding_period(df, date, "month", complete = FALSE)              # keep partial windows
analysis(sp$splits[[1]])                                         # training rows of split 1

What sliding_period() does

sliding_period() partitions a time-ordered data frame into resamples whose boundaries follow calendar units. It belongs to the rsample package and complements sliding_window() (row-based) and sliding_index() (column-value-based). The key argument is period, which accepts "year", "quarter", "month", "week", "day", "hour", "minute", or "second". Internally the function uses lubridate::floor_date() to bucket the index column into period bins, then slides over those bins.

The function returns a tibble with a splits list-column and an id column. Each row is one rsplit object whose analysis and assessment indices you extract with analysis() and assessment(). Unlike row-based resamplers, sliding_period() produces splits of variable row count: a month with 31 days yields more rows than a month with 28. This matters when daily traffic, sales, or measurement frequency drifts over time, because every fold still represents the same calendar exposure.

Syntax and arguments

The signature has one data argument, one index column, one period unit, and seven window knobs.

The arguments that matter most:

• index: an unquoted column name holding a Date or POSIXct value. The column controls bin assignment but does not have to be sorted.
• period: the calendar unit each bin represents. "month" is the most common choice for business reporting.
• lookback: the number of complete periods before the current position included in the analysis window. lookback = 11 with period = "month" gives a 12-month rolling training window. lookback = Inf produces an expanding window.
• assess_start and assess_stop: offsets ahead of the current position that define the assessment window in period units. Defaults give a one-period, one-step-ahead test.
• every: how many periods are grouped into a single bin. period = "week", every = 2 produces fortnightly bins.
• complete: TRUE drops splits whose lookback would extend before the first period; set FALSE to keep them.

sliding_period() examples

Monthly rolling 12-month backtest

Use lookback = 11 with period = "month" to train on the last 12 months and test on the next. This is the canonical setup for one-month-ahead forecasting of monthly aggregates.

The splits list-column shows row counts that vary by month length. Notice the analysis window holds 365 or 366 rows (one calendar year) and the assessment window holds 28 to 31 rows (one month).

Three-month forecast horizon

Set assess_stop = 3 to evaluate a three-month-ahead forecast. The assessment block expands from one period to three while training stays the same.

Each split now trains on 12 months and assesses on the following quarter, mirroring how operations teams review forecasts on a quarterly cadence.

Weekly resamples with fortnightly bins

Combine period = "week" with every = 2 to bin the data into two-week blocks. This works when daily data is too noisy for week-by-week comparison but monthly aggregation hides intra-month structure.

Expanding training window

Pass lookback = Inf to keep all earlier periods in the analysis window. This is the calendar-period equivalent of rolling_origin(cumulative = TRUE).

The first split trains on January, the tenth trains on January through October. Use this when older history still carries signal and discarding it would waste data.

sliding_period() vs sliding_window() vs rolling_origin()

Pick sliding_period() when calendar boundaries matter, sliding_window() for fixed row counts, and rolling_origin() for the classic expanding-origin backtest. The table below summarizes the trade-offs.

Common pitfalls

Three mistakes show up in nearly every first-time use.

• Unsorted index column without realizing it. sliding_period() does not error on unsorted data; it bins rows by date but the resulting analysis() and assessment() tibbles inherit the original row order. Always arrange(date) before resampling if you plan to feed splits into a model that assumes ordered input.
• Confusing lookback with row count. lookback = 12 does NOT mean 12 rows. With period = "month" it means 12 months of rows, which may be hundreds. Read lookback as "periods", not "rows".
• Forgetting complete = TRUE drops early splits. The first complete window cannot start before period lookback + 1, so the first lookback periods produce no splits by default. If you need them, set complete = FALSE and handle the shorter windows downstream.

Try it yourself

Related rsample functions

These resamplers either pair with sliding_period() or replace it for adjacent use cases.

• sliding_window(): fixed row count instead of calendar period.
• rolling_origin(): classic expanding or sliding row-based backtest.
• sliding_index(): slide by values of an arbitrary index column.
• vfold_cv(): standard k-fold cross-validation when ordering does not matter.
• initial_time_split(): single chronological train and test split.

See the official rsample reference for the full argument matrix.

FAQ

What is the difference between sliding_period() and sliding_window() in rsample?

sliding_period() partitions data by calendar units (months, weeks, days) using a date index column, so window sizes adapt to the actual row density of each period. sliding_window() partitions by fixed row count, so every window contains exactly the same number of rows. Use sliding_period() when business reporting follows calendar boundaries; use sliding_window() when data is evenly sampled and identical training sizes matter for model comparison.

Can sliding_period() handle daily data with missing dates?

Yes. The function bins each row by its date value, so missing dates simply produce smaller analysis windows for the affected periods. The window still spans the same calendar range, but with fewer rows. This is one of the main reasons to prefer sliding_period() over sliding_window() for messy real-world time series with holidays, sensor downtime, or weekend gaps.

How do I make an expanding window with sliding_period()?

Set lookback = Inf. Every split's analysis window then includes all periods up to and including the current one. The first split trains on the first complete period, the second on the first two, and so on. This is the calendar-period equivalent of rolling_origin(cumulative = TRUE), and it is the right choice when older history still carries predictive signal.

Does sliding_period() require sorted data?

No, but you should sort anyway. The bin assignment uses the date column directly, so unsorted input produces correct splits. However, the rows within each analysis() and assessment() tibble inherit the input order, which can confuse downstream models that assume chronological ordering. Run arrange(date) before resampling as a safe default.

What period units does sliding_period() accept?

The supported units are "year", "quarter", "month", "week", "day", "hour", "minute", and "second". These map to the period names accepted by lubridate::floor_date(). Combine the unit with the every argument to get multi-unit bins like "every = 2" weeks (fortnightly) or "every = 6" hours (quarter-day).