Zombie Steve and the Hockey Stick
September 29, 2014toto said… My understanding is that "short-centering" was actually a bona-fide bug. If you’re computing a covariance matrix by simply multiplying the data with its own inverse, "short-centering" the data around the recent past will produce "fake covariance" of all series with the hockey-stickish series (with a negative sign, but that’s irrelevant).The actual PCs being computed are the same, but the ‘hockey-stickish’ ones now seem to explain much more of the variance – spuriously so.Unfortunately the auditor-in-chief proceeded to introduce a bug of his own, by keeping the same number of PCs as Mann after full-centering the matrix, even though by then you need to include more PCs to capture sufficient variance
Read more here:
Zombie Steve and the Hockey Stick