Purged cross-validation is a time-series validation method that removes training observations whose label periods overlap the test set in time.

Purged cross-validation is a time-series validation method that removes training observations whose label periods overlap the test set in time. The purpose is to prevent information from the test labels from leaking into the training sample through overlapping event horizons, something standard k-fold cross-validation does not account for in sequential financial data.

In ordinary k-fold cross-validation, observations are split into folds without regard to how labels are formed over time. That approach can work for independent and identically distributed data, but it is not reliable when each label depends on a forward return window or another time interval that extends beyond the observation timestamp. In that setting, a training example can contain information from a period that also contributes to a test label. The split looks clean by row index, but it is not clean in time.

Purged cross-validation addresses this by removing from the training fold any observations whose label spans intersect the test interval. The mechanism is simple. First, define the start and end time of each labeled observation. Next, choose a test fold. Then exclude from the training set every observation whose label end time reaches into the test window or whose label interval otherwise overlaps the test labels. The remaining training data no longer shares label-time information with the test fold.

This matters because the leakage in plain k-fold is forward‑looking. A model can appear to generalize out of sample when it has indirectly seen information from the same future period through overlapping labels in the training set. Purging closes that path. It makes the validation procedure better aligned with how a strategy would be trained and evaluated in live use, where future label information is unavailable at fit time.

Purged cross‑validation is a validation design choice for path‑dependent labels rather than a guaranteed numerical improvement.

Covered in depth in the Strategy validation & overfitting pillar hub.