When you fit a regression model, the fit statistics you compute on the data used for fitting (R-squared, residual standard error, training error) are optimistic. The model has already seen those points, so it can partly memorize their noise. Cross-validation is the family of techniques that answers a different and more useful engineering question: how accurately would this model predict a new observation it has never seen?
This topic covers the main approaches in increasing sophistication. The holdout method splits the data once into a training set and a test set. K-fold cross-validation rotates the role of the test set across k partitions so every observation is predicted exactly once out-of-sample, then averages the error. Leave-one-out cross-validation (LOOCV) is the extreme case where k equals the sample size. Each choice trades off bias, variance, and computational cost, and each is appropriate in different situations — a distinction you are expected to be able to make.
While watching, focus on three things: the mechanics of how data are split and reused, why the estimated test error behaves differently than the training error, and how the resulting cross-validation curve is read when comparing candidate models of different complexity. That last point connects directly to model refinement — choosing which predictors or functional forms to keep is a decision you can justify with cross-validated error rather than in-sample fit alone.
Trevor Hastie introduces cross-validation as a way to estimate test error when a separate test set is not available, contrasting it with training error and explaining why training error understates prediction error. Based on the title and the Statistical Learning curriculum, this opening lecture should cover the validation-set (holdout) approach and its drawbacks, including the variability that comes from a single arbitrary split.
It is the authoritative first lecture on the subject from Stanford Online, taught by one of the authors of the standard text, and it maps directly onto the "Why Cross-Validation Is Needed" and "Holdout Method" subtopics.
This follow-on lecture moves from the single holdout split to k-fold cross-validation, describing how the data are partitioned into k folds and how the fold-level errors are combined into one estimate of test error. Given the title and sequence, expect discussion of LOOCV as the special case k equals n, the common choices of k such as 5 or 10, and the bias-variance and computational trade-offs among them.
It covers the K-Fold and LOOCV subtopics in one place and gives you exactly the procedural comparison you need to distinguish the methods and justify a choice of k.
A compact, under-nine-minute overview that, per its description, walks through the basics of training and test data and then explains why cross-validation is needed before presenting leave-one-out and k-fold schemes. It is pitched as part of an introductory statistical learning course, so expect a straightforward, example-led treatment rather than heavy theory.
It is a fast, accessible summary that reinforces the same vocabulary as the Stanford lectures, making it useful as a preview before them or as a review afterward if the longer lectures move quickly.
These three videos concentrate on MO3: together they define the holdout split, k-fold cross-validation, and LOOCV, lay out the procedure for each, and compare them on variance, bias, and computational cost so you can state when each is appropriate. They also support MO5 indirectly, since the cross-validation error curve is the standard evidence base for deciding whether to drop a variable, change a functional form, or accept a simpler model — refinement decisions justified by out-of-sample performance rather than in-sample fit. The residual-diagnostic outcomes (MO1, MO2, MO4) are addressed by other materials in this module; use cross-validation alongside those plots, not in place of them, since a model can generalize reasonably well and still violate assumptions in ways residual analysis will reveal.
These videos were chosen from YouTube on October 2, 2026. Each has captions; the player turns them on. Videos on YouTube can be changed or removed — if one no longer plays, tell your instructor. If a player does not appear, use its Watch on YouTube link.