LoadingLoading
A k-fold cross-validation variant for financial time series that purges training labels overlapping the test set and applies a post-test embargo to prevent leakage.
A k-fold cross-validation variant for financial time series that purges training labels overlapping the test set and applies a post-test embargo to prevent leakage.
Proposed by Marcos López de Prado in Advances in Financial Machine Learning (2018, Chapter 7), purged k-fold CV fixes the leakage that standard k-fold introduces in financial data. Because labels such as triple-barrier outcomes span a holding horizon, a training observation formed before the test set can still overlap it in label time, and serial correlation leaks information across adjacent samples. The procedure splits data into K folds, purges every training observation whose label interval [t0, t1] overlaps any test-set interval, then applies an embargo that drops a small fraction of training samples immediately after each test fold before fitting the model.
Standard k-fold assumes i.i.d. data; financial labels violate this through overlapping horizons and serial correlation, so naive CV overstates out-of-sample skill.
Purging removes training rows whose label window touches the test window, even when the feature timestamp t0 sits outside the test fold.
The embargo drops ~1% of samples (or roughly one max holding period) after each test fold to absorb residual autocorrelation and event clustering.
Practical defaults are K = 5-10 folds with embargo sized to the longest label horizon; larger K sharpens the estimate but raises purge cost and variance.
Use purged k-fold for model and hyperparameter selection, walk-forward for production-behavior simulation, and CPCV when you need a distribution of backtest paths and a probability of backtest overfitting.
Pairs naturally with triple-barrier labeling, which is the main source of label overlap that makes purging necessary.
Research methodology is the set of practices that distinguish a rigorous quant research process from a hopeful one. The central topics are backtest audit, multiple testing correction, purged and combinatorial cross-validation, walk-forward optimisation, the triple-barrier labelling method, and the deflated Sharpe ratio. The dominant practitioner reference is López de Prado (2018).
Financial machine learning is the application of supervised, unsupervised, and reinforcement learning methods to financial prediction, classification, and decision problems. The defining methodological constraint is that financial data are serially correlated, not independently and identically distributed, which means that the standard machine learning toolkit must be substantially adapted. The dominant practitioner reference is López de Prado (2018).