Explain precisely why standard k-fold cross-validation is optimistic on a time series of overlapping return labels. Name the two distinct leakage channels and say which one purging fixes and which…

Question

Explain precisely why standard k-fold cross-validation is optimistic on a time series of overlapping return labels. Name the two distinct leakage channels and say which one purging fixes and which one embargoing fixes.

Accepted Answer

Random k-fold leaks through label overlap, where a training label spans a period that also determines a test label, and through serial correlation of features immediately after the test block, and purging removes the overlapping training rows while embargoing removes a buffer of rows following the test block. A label like the five-day forward return at time t depends on prices from t to t + 5, so any training row with an index inside that window shares the same realised outcome and the model can memorise it rather than predict it. Purging deletes every training observation whose label interval intersects any test label interval, which is a function of the label horizon, not of the fold size. The second channel is different: even with no overlap, a training row shortly after the test block has features built from an autocorrelated series and a label driven by the same regime, so information flows backwards into the test estimate. The embargo drops a fixed span of rows after each test block, typically a small percentage of the sample, to break that. Both are needed, and neither addresses the third problem, that shuffling folds trains on the future to predict the past, which walk-forward evaluation handles by construction.