K-Fold Cross-Validation
Splitting data into K parts and training K times, each run validating on a different part, so every row is validated exactly once.
The point is that it returns a spread as well as a mean, and the spread is the half people drop from the report. At a fold standard deviation of 2 points, a 5-fold mean has a standard error of , so a 0.3-point difference between two models means nothing.
Five or ten folds is standard. Small K is pessimistic because each model sees less data; large K has lower bias and more correlated folds. Leave-one-out is the K = n limit, with the lowest bias and a noisy estimate.
Two variants are mandatory rather than optional: grouped folds when rows share an entity, and a time-series split — never shuffled, training always earlier than validation — when rows are ordered in time. And every fitted step refits inside the fold, which is what a pipeline passed to the cross-validator gives you.