[kqr]

No, the post is doing cross-validation to test predictive power directly. The error will not decompose as neatly then.

[fluidcruft]

Why would they do that and where do you see evidence they did?

[kqr]

Because it's a direct way to measure predictive power, and it says so: "We’ll use leave-one-out cross-validation"