Stratified Cross-Validation Calculator

Calculated report

Cross-validation summary

Review fold balance, stability, risks, and exportable results.

Distribution

Dataset and holdout allocation

Counts and percentages show the data available before fold creation.

Class	Total	Original %	Holdout test	CV pool	CV pool %

Fold details

Training and validation distribution

Each row represents one fold or stratified shuffle split.

Show assigned sample indexes

Comparison

Stratified versus standard K-Fold

Lower deviation means validation folds follow the original class ratios more closely.

Method	Mean class-ratio deviation	Maximum deviation	Empty class-fold cells	Interpretation

Performance

Aggregated metric results

Statistics use the fold-level values entered below.

Metric	Count	Mean	Median	Minimum	Maximum	Range	Std. deviation	Variance	Std. error	Confidence interval	CV	Best fold	Worst fold

Class evaluation

Combined confusion matrix and class metrics

Precision, recall, and F1-score are calculated from optional confusion counts.

Visual review

Charts

Use the charts to inspect balance, deviation, performance, and repetition stability.

Original class distribution

Validation classes by fold

Training versus validation size

Class-ratio deviation

Fold metric performance

Metric confidence intervals

Repetition stability

Portable output

Merged report data

Copy this text into notes, experiments, or model documentation.

Step 1

Choose the dataset source

Enter class frequencies manually or import labels from CSV data.

Input source

Manual value type

Total samples

Required when percentages are entered.

Class frequencies

Step 2

Configure cross-validation

Choose the splitting method, fold count, repeat count, and holdout policy.

Validation method

Number of folds

Repetitions or shuffle splits

Shuffle before splitting

Random seed

Separate holdout test set

Displayed decimal places

Distribution display

Confidence level

Step 3

Add fold-level performance metrics

Paste a header row followed by one row per fold. Missing values are allowed.

AccuracyPrecisionRecallSpecificityF1Balanced accuracyROC-AUCLog lossMCCCohen kappaCustom

Metric table

Use commas, tabs, semicolons, or spaces. Values may be decimals or percentages.

Step 4

Add optional confusion counts

Use one line per cell: fold, actual class, predicted class, count.

Confusion matrix rows

Fold names may be numbers or labels. Class names must match the dataset labels.

Examples

Load example data

Use a preset, then adjust any value before calculating.

Formula used

Core calculations

These formulas summarise fold sizing and score variability.

Validation samples per fold ≈ N ÷ K

N is the cross-validation sample count. K is the fold count.

Class samples per fold ≈ Nc ÷ K

Nc is the sample count for one class.

Mean score = Σsi ÷ K

si is the score from fold i.

Sample standard deviation = √[Σ(si − mean)² ÷ (K − 1)]

A larger value indicates less stable validation performance.

Standard error = standard deviation ÷ √K

Confidence intervals use the selected confidence level.

How to use

Calculation steps

Enter class counts, percentages, or CSV labels.
Choose the validation method and fold count.
Set repeats, shuffling, seed, and holdout percentage.
Optionally paste fold metrics and confusion counts.
Calculate, inspect warnings, compare methods, and export.

Example data

Common classification scenarios

Scenario	Class counts	Suggested folds	Main concern
Balanced binary	50, 50	5 or 10	General stability
Imbalanced binary	24, 476	5	Minority representation
Three classes	120, 75, 45	5	Multi-class balance
Small minority	4, 196	At most 4	Empty minority folds
Repeated validation	90, 60, 30	5 repeated	Variance estimation
Ten-fold example	130, 170	10	Smaller validation folds

Learning guide

Important cross-validation practices

Stratification preserves class proportions across validation folds. It is especially useful for imbalanced classification datasets. It does not repair weak labels or leakage.

Fit preprocessing inside every training fold only. This includes scaling, selection, encoding, and resampling. Applying them globally can inflate validation scores.

Group-aware data needs grouped stratification methods. Multilabel targets need iterative multilabel stratification. Ordinary stratification may separate related samples incorrectly.

Leave-one-out classification can produce unstable class coverage. Repeated stratification usually estimates variability more clearly. Use an untouched test set for final reporting.

Frequently asked questions

Stratified cross-validation questions

What does stratified cross-validation preserve?

It keeps each fold close to the dataset’s class proportions.

How many folds should I use?

Five or ten folds are common. The smallest class limits the safe maximum.

What happens when a class has fewer samples than folds?

Some validation folds cannot contain that class. Reduce the fold count.

Should I shuffle before splitting?

Usually yes, unless the data order carries time or sequence meaning.

Why use repeated stratified validation?

Repeats reduce dependence on one random split and expose score variance.

Can stratification prevent data leakage?

No. Pipelines and group-aware splitting are still required.

Should oversampling happen before cross-validation?

No. Apply oversampling only inside each training fold.

Can I use this calculator for regression?

Direct class stratification is for classification. Regression requires binned targets or other strategies.

What indicates unstable model performance?

Large standard deviation, wide confidence intervals, and inconsistent fold scores indicate instability.