Label overlap occurs when the future-information window used to define one observation's target intersects the information window used by another observation.
A training row can sit outside the nominal validation dates and still leak validation information if its target reaches into the validation period.
Chronological separation alone may not stop leakage
Suppose each label uses the next five rows of returns. A training observation immediately before the validation fold can still consume returns that belong to the validation interval.
Purging removes those directly overlapping training labels. An embargo period can then add extra temporal separation after direct overlap ends.
That is different from a holdout set, which describes reserved evaluation data at a higher level.
Use the Purged Cross-Validation calculator for the row-level overlap geometry.
Label horizon must match the actual research target
Overlap depends on how the label is constructed. A one-day forward return, a 20-day barrier event, and a multi-month outcome create different information intervals.
Purging the wrong horizon can leave leakage in place or discard more training data than necessary.
Sources: RiskLab AI, Cross-Validation in Finance and Grizzly Bulls Purged Cross-Validation.
Continue Research
Continue from the concept into the Grizzly Bulls research surface that best matches the next question. These links are research continuations, not recommendations or required steps.
Purge overlapping labels
Use explicit forward label windows to identify training rows that still consume validation information.
Explore more topics in the Financial Research Encyclopedia.