Purging and Embargoing to Prevent Leakage in Financial Cross-Validation

Summary

The article explains why ordinary cross-validation can mislead when financial features and labels depend on overlapping time periods. Embargoing removes observations near a test fold when their feature lookbacks could draw on information from that fold. Purging removes training observations whose label event periods overlap test-fold trade periods, such as when a stop or target is reached after entry. The article recommends applying both safeguards when evaluating strategies, including rule-based systems, to reduce leakage and overfitting.

It also describes combinatorial purged cross-validation as a way to construct multiple backtest paths from historical data. A SPY example reports a high full-history Sharpe ratio alongside a very large maximum drawdown, illustrating why a single headline metric can be misleading. The example and reported outputs are limited; they do not establish that the strategy is robust or suitable for live trading. The article stresses disciplined evaluation across folds and warns that repeatedly tuning against historical results can select random patterns.

Key ideas

• Feature lookbacks can leak information across chronological folds, so embargo affected observations near fold boundaries.
• Purge training examples when their label event periods overlap the test fold’s trade periods.
• Use cross-validation to assess both parameter stability and out-of-sample strategy performance.
• Combinatorial purged cross-validation can generate multiple backtest paths from one historical series.
• A strong Sharpe ratio can coexist with severe drawdowns, so evaluate more than one performance measure.

Tags

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.