Cross-Validation

Cross-validation is the evaluation protocol that splits available data into multiple folds, trains the model on most folds, tests on the remaining fold, and averages the scores — producing a more reliable estimate of out-of-sample performance than a single train/test split. **k-fold cross-validation** (k=5 or k=10 is standard) rotates through every fold as the held-out set. **Stratified k-fold** preserves class proportions in each fold — mandatory for imbalanced classification. **Leave-one-out** cross-validation is the k=n extreme — expensive but low-variance; used on very small datasets.

The technique exists to fight a specific failure mode: reporting model quality from a single 80/20 split gives you one number with unknown variance. That number can vary by several percentage points across different random splits of the same data. Cross-validation reports a mean and a standard deviation, which is what a business decision actually needs. Two models whose single-split scores differ by 1% might be indistinguishable under cross-validation, in which case the cheaper-to-run or easier-to-explain one should win.

For time-series and sequential data, standard k-fold violates causality — test folds contain future information that informs past training folds. **TimeSeriesSplit** (expanding or rolling window) fixes this by always putting the test fold chronologically after the training folds. For grouped data (multiple samples per customer, patient, or session), **GroupKFold** prevents leakage by keeping all samples from a group in the same fold. Getting the fold construction wrong is one of the most common silent ways real ML projects ship inflated metrics that never replicate in production.

For APAC mid-market, cross-validation is the cheap insurance that catches bad data splits before they bite. The rule to enforce: no model ships to production on the strength of a single split. If you cannot afford k=5, at least report across three different random seeds. The extra compute cost is small; the reputational cost of a model that looked great in the lab and failed in production is not.