An embargo period is a buffer of observations removed after each test window in a validation scheme.

An embargo period is a buffer of observations removed after each test window in a validation scheme. The purpose is to keep nearby data from crossing the boundary between test and train segments when returns, labels, features, or events remain serially dependent across time. In strategy validation, this buffer reduces the chance that information tied to the test window reappears in the training window through overlap or persistence in the data generating process.

The mechanism is straightforward. If observations close in time are correlated, then a train sample placed immediately after a test sample can still contain information linked to the test period. This can happen when labels span multiple timestamps, when features are built from rolling windows, or when market effects decay gradually rather than ending at the boundary. By dropping a short interval after the test window, the validation procedure widens the separation between the two samples. That separation lowers leakage risk and makes the evaluation closer to a genuinely unseen forecast setting.

Sonar’s strategy validation material presents embargoing as part of a leakage aware validation framework for financial machine learning and backtesting. The discussion is tied to the broader need to control overlap, dependence, and look ahead effects when estimating model quality on time ordered data. Within that framework, embargoing is used alongside purging to prevent contamination between folds when labels or events extend through time. This supports the general claim that an embargo can improve the reliability of strategy validation by making train and test partitions more independent in practice.

An embargo period is designed to reduce boundary leakage arising from serial dependence.

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