A Practitioner's Guide to Cross-Validation in Finance

Standard k-fold cross-validation can be dangerously misleading when applied to financial time-series data. Here's why, and what to do about it.

The Problem with Standard CV

Financial data has three properties that violate the assumptions of standard cross-validation:

Serial correlation -- Today's returns are correlated with yesterday's returns

Non-stationarity -- The data distribution changes over time (regime shifts)

Information leakage -- Features computed from a rolling window can leak future information

Better Alternatives

Purged Walk-Forward CV

from sklearn.model_selection import TimeSeriesSplit

def purged_walk_forward(X, y, n_splits=5, embargo_pct=0.01):
"""
Walk-forward CV with purging and embargo.
Purging removes samples near the train/test boundary.
Embargo prevents using recently trained-on data.
"""
tscv = TimeSeriesSplit(n_splits=n_splits)
embargo_size = int(len(X) * embargo_pct)


for train_idx, test_idx in tscv.split(X):
# Purge: remove overlap
train_idx = train_idx[:-embargo_size]
yield train_idx, test_idx

Combinatorial Purged CV (CPCV)

Introduced by Marcos Lopez de Prado, CPCV generates multiple train/test paths through the data, providing more reliable backtests with fewer data points.

Key Takeaways

• Never use random k-fold CV on time-series data
• Always purge samples near train/test boundaries
• Apply an embargo period proportional to your feature lookback
• Use multiple paths through the data to reduce variance of your performance estimate

Financial ML requires financial ML techniques. Don't trust generic tools blindly.