Purged K-Fold Cross-Validation: Eliminating Look-Ahead Leakage

Standard $k$-fold cross-validation partitions the dataset into $k$ folds, trains on $k-1$, and tests on 1, cycling through all folds. In finance this fails: financial returns are serially correlated and overlapping labels (returns spanning multi-day windows) create information leakage from the testing set into the training set. A classifier trained on observations that inadvertently share information with the test set will overestimate its out-of-sample performance — even if the test fold was never explicitly seen during training.

Purged $k$-fold CV fixes this with two operations: purging (removing training observations whose labels overlap with any testing label) and embargoing (removing training observations that immediately follow the test set, where serial correlation may carry test information backward). Together, these ensure the training set contains no information that could not have been known at the start of the testing period.

Learning objectives. By the end of this note you should be able to: