Learning goal: changing K affects training size, estimation dependence, and cost. Five or ten folds is a common compromise, not a universal optimum. An important failure is placing fold boundaries across dependent observations inappropriately.

CV variability and scaled bias versus K; cross: single-holdout reference.

Random-fold and forward-chaining scores versus temporal correlation; dashed: the defined reference risk.

Reference risk = — Single-holdout standard deviation = — 5-fold std = — 10-fold std = — LOO std = —

Observe: this setup's variability changes quickly at small K, then shows limited further improvement over part of the range. The lower plot addresses a different issue: strong temporal correlation can make random-fold scores optimistic for the intended future-prediction task. Forward chaining better matches that task here, but needs suitable temporal assumptions and a matching risk definition.