Why k-fold cross-validation lies on time series

Question

You validate a return-prediction model with standard $10$ -fold cross-validation (rows shuffled into folds). It scores well, but live trading disappoints.

What is wrong with plain k-fold here, and what should you use instead?

Accepted Answer

Random k-fold assumes rows are independent and interchangeable. Financial time series violate both, in two damaging ways. 1. It trains on the future. With shuffled folds, some training rows come from dates after the test rows. The model gets to peek past the prediction point, an unrealizable advantage that inflates the score. Live, there is no future to look at. 2. Adjacent rows leak across folds. Returns and especially features (moving averages, rolling vols) are autocorrelated, so a test row and its neighbor in the training set are nearly the same observation. The model effectively sees its…

Why k-fold cross-validation lies on time series

Your answer

More Statistics questions