Purged Walk-Forward (with embargo)
Education only · our voice · free public data
Definition
A time-ordered validation scheme that repeatedly trains on a past window and tests on the immediately following window, while PURGING training samples whose labels overlap the test period and EMBARGOING a buffer of samples right after each test set to block serial-correlation leakage.
How to read it
Walk-forward mimics live deployment: you only ever use the past to predict the future, and you roll the window forward through time. The purge and embargo are what make it rigorous. Purging removes any training label that was constructed using data inside the test window (common when labels use forward-looking horizons); the embargo drops a few observations after the test set so autocorrelation cannot bleed test information back into subsequent training. Without them, an ML strategy can post a spotless backtest that is pure leakage.
How practitioners use it
Used as context among multiple indicators — never as a standalone signal to act.
Less common professional uses
Purging targets LABEL overlap, not feature overlap: if a label at time t is defined over [t, t+h], any training observation whose label window intersects the test window must be dropped, otherwise the model has literally seen the answer. Embargo defends against the fact that returns just after a test block remain correlated with it; the correct embargo can be asymmetric and must exceed the maximum feature-lookback plus label-horizon to fully decorrelate. Combinatorial Purged Cross-Validation generalizes walk-forward by testing on multiple non-contiguous groups at once, yielding many backtest PATHS and hence a sampling distribution of Sharpe - this is what feeds a credible Deflated Sharpe. Walk-forward with periodic re-optimization is itself a multiple-testing engine; each re-fit is a trial, so the aggregate result still needs data-snooping deflation.
Sources & provenance
Lopez de Prado (2018), 'Advances in Financial Machine Learning', Ch. 7 (Purged CV) & Ch. 11-12; Pardo (2008), 'The Evaluation and Optimization of Trading Strategies'
This page is educational content published by Pachira Aquatica Global LLC. It is not investment advice and not a recommendation.