rsample mc_cv() in R: Monte Carlo Cross-Validation Splits

The rsample mc_cv() function in R builds Monte Carlo cross-validation splits, drawing many independent random train/test partitions so you can average a model's performance across resamples without committing to a fixed v-fold structure.

mc_cv(df)                                # default: 25 splits, prop = 0.75
mc_cv(df, prop = 0.8)                    # 80/20 random splits
mc_cv(df, times = 100)                   # 100 Monte Carlo splits
mc_cv(df, prop = 0.7, times = 50)        # 50 splits at 70/30
mc_cv(df, strata = y)                    # class-balanced splits
analysis(mcs$splits[[1]])                # training rows of split 1
assessment(mcs$splits[[1]])              # held-out rows of split 1
set.seed(123); mc_cv(df, times = 25)     # reproducible splits

What mc_cv() does

mc_cv() draws a sequence of independent random train/test splits from a data frame. It belongs to the rsample package, the resampling engine of the tidymodels ecosystem. Each call to mc_cv() returns a tibble of rsplit objects in a splits list-column, where every row is one Monte Carlo replicate with its own analysis (training) portion and assessment (held-out) portion.

The defining feature is independence. Unlike vfold_cv(), which partitions every row into exactly one held-out fold, mc_cv() resamples freely. A given row may sit in the assessment set of several splits or of none at all. That property makes mc_cv() useful when you want many resamples (50, 100, or more) without inflating the cost of a high-v fold scheme, and when the natural unit of evaluation is a random hold-out rather than an exhaustive partition.

Syntax and arguments

The signature has one required argument and four tuning knobs.

The arguments that matter in practice:

• data: the data frame or tibble to resample.
• prop: proportion of rows assigned to the analysis (training) set in each split. The default 0.75 gives 75/25 splits.
• times: how many independent splits to draw. The default 25 is a reasonable starting point; raise to 50 or 100 on small datasets.
• strata: a column name. Each split is drawn so the class proportions of strata are preserved in both analysis and assessment sets.
• breaks: when strata is numeric, the number of quantile bins used for stratification.

mc_cv() examples

Default Monte Carlo splits

Call mc_cv() on a data frame and print the result to see the split structure. The defaults produce 25 splits at a 75/25 ratio.

Extract analysis and assessment sets

Each split exposes an analysis() set for fitting and an assessment() set for scoring. These helpers return ordinary data frames you can pass to any modeling function.

Change the train/test proportion

Pass prop to control how much of the data goes into the analysis set on each split. Raising prop produces smaller assessment sets and lower-variance training, useful when the data is scarce and you want to keep most of it for fitting.

Increase the number of resamples

Set times higher to average over more random splits. A common pattern is 50 or 100 resamples on small datasets where a single 10-fold pass is noisy.

Stratify by the outcome

Use strata to keep a class or numeric distribution balanced across every split. Without stratification a random split can under-represent a rare class. Here a cylinder-balanced split preserves the 4/6/8-cylinder mix in both portions.

mc_cv() vs other resampling functions

mc_cv() draws repeated random hold-outs; the alternatives have different sampling structures. Choose based on the resampling estimator you need.

A tidymodels workflow can plug mc_cv() directly into fit_resamples() or tune_grid() wherever a vfold_cv() object would go. The resamples object is interchangeable.

Common pitfalls

Three mistakes account for most mc_cv() bugs.

• Forgetting set.seed(). mc_cv() shuffles rows on every call. Without set.seed() before the call, runs are not reproducible and any cross-validated metric drifts between executions. Place the seed in the same script, immediately before the resample.
• Using mc_cv() on time-ordered data. Random splits ignore chronology, so a forecasting model can train on the future and test on the past. Use rolling_origin() instead for time series; reserve mc_cv() for cross-sectional data.
• Confusing times with prop. times controls how many splits are drawn, not how big each one is. To take 100 resamples of size 30/70 use mc_cv(df, prop = 0.3, times = 100), not times = 30, prop = 100.

Try it yourself

Related rsample functions

mc_cv() draws the splits; these companions consume and extend them.

• analysis() and assessment(): pull the training and held-out data frames from a split object.
• initial_split(): carve off a final testing set before applying mc_cv() to the training portion.
• vfold_cv(): build a partitioned v-fold object when every row must be held out exactly once.
• bootstraps(): draw with-replacement resamples when variance estimates are the goal.
• group_vfold_cv(): keep all rows of a group together; mc_cv() has no group-aware analogue.

FAQ

What is the default number of splits in mc_cv()?

The default times argument is 25, which builds 25 Monte Carlo splits at a 75/25 ratio. Each split is independent, so a row can be in the assessment set of several splits or none. Twenty-five splits is a reasonable starting point for medium-sized datasets. On small data, raise it to 50 or 100 to reduce the variance of the averaged performance metric. The cost scales linearly with times.

How is mc_cv() different from vfold_cv()?

vfold_cv() partitions the data: every row sits in exactly one held-out fold across v folds, so v controls both the number of folds and the held-out size. mc_cv() resamples freely: it draws each split at random with the chosen prop, and the same row can appear in the assessment set of many splits or none. Use vfold_cv() for standard k-fold cross-validation and mc_cv() when you want many independent random hold-outs.

How do I make mc_cv() reproducible?

Call set.seed() with any integer immediately before mc_cv(). The function randomises rows on every call, so a fixed seed guarantees identical splits across runs. Keep the seed and the resample call in the same script, and rerun with the same seed plus the same R version to reproduce the exact assignments. The splits column is stable under that combination.

Can mc_cv() be used for time series data?

No. mc_cv() ignores row order and draws random hold-outs, which lets a forecasting model train on future observations and test on past ones. That leakage destroys the validity of any time-series performance estimate. Use rolling_origin(), sliding_window(), or sliding_period() from rsample instead. Those functions respect chronology and produce resamples that mimic real forecasting evaluation.

Does mc_cv() work with tune_grid() and fit_resamples()?

Yes. The object returned by mc_cv() carries the same rset class as vfold_cv() output, so tidymodels tuning and evaluation functions accept it without modification. Pass it as the resamples argument to fit_resamples() for a single model, or to tune_grid() when tuning hyperparameters. The performance metrics are then averaged across the times Monte Carlo splits rather than across v folds.