Purged Cross-validation

What is purged cross-validation?

Purged cross-validation is a modification of standard k-fold cross-validation for financial machine learning, designed to stop information leaking between training and test folds through overlapping labels. Popularized by Marcos Lopez de Prado in Advances in Financial Machine Learning (2018), it adds two operations, purging and embargo: purging deletes from the training set any observation whose label window overlaps in time with a test observation's label window, and the embargo additionally drops a buffer of training observations immediately after each test block to absorb serial correlation that outlives the labels.

The problem it fixes is subtle but severe. Financial labels are usually computed over a forward window, for example the return over the next 20 bars or the outcome of a triple-barrier exit. When ordinary k-fold assigns neighboring observations to different folds, a training sample's label can be built from the very same price path a test sample is judged on. The model is then partly graded on information it saw in training, cross-validated scores inflate, and hyperparameter tuning confidently selects overfit models. Because returns and volatility are serially correlated, even non-overlapping neighbors leak, which is what the embargo addresses.

Traders and quant researchers care because leakage of this kind is a dominant reason machine-learning strategies ace validation and fail live. The combinatorial extension, CPCV (combinatorial purged cross-validation), applies purging and embargo across many train/test block combinations, producing multiple backtest paths instead of one and thereby a distribution of performance rather than a single, easily overfit number.

How traders use it

• As the default validation scheme when training ML models on financial labels: define each sample's label interval explicitly, purge overlapping training samples around every test fold, and add an embargo of a small percentage of the data after each test block.
• For hyperparameter tuning, where leakage does the most damage: tuning against leaky folds systematically selects complexity that memorizes shared price paths, so purged folds are used inside the tuning loop, not just for the final score. This is part of broader train/validation discipline.
• Via CPCV when a distribution of outcomes is wanted: enumerating many purged train/test combinations yields multiple out-of-sample paths, whose spread feeds overfitting diagnostics in the same spirit as the probability of backtest overfitting.
• With label design done first: purging requires knowing exactly when each label's information begins and ends, so a precise label definition and prediction horizon is a prerequisite rather than an afterthought.
• Within its limits: purging removes leakage through labels and nearby autocorrelation, but it does not defend against feature engineering that already used future data, nor against overfitting from testing many model variants, which still needs its own correction.

Purged cross-validation vs related concepts