What Is Embargo?
In time-series validation, an embargo is a buffer interval after a test fold during which observations are excluded from training. It reduces information leakage when nearby observations share labels, lookback windows, or serially dependent features with the test period.
The embargo length should be chosen from the experiment’s label horizon and data-dependence structure, not assumed to be universal. It is commonly paired with purged cross-validation, which removes observations whose label windows overlap the test fold itself.
See also
Research and literature
Marcos López de Prado, Advances in Financial Machine Learning, Wiley, 2018.