Glossary
Cross-validation
Estimating how well a model works on unseen data by repeatedly fitting on part of the sample and testing on the rest.
In k-fold cross-validation the sample is split into k folds. Each fold in turn is the test set and the others train the model. Averaging the k scores estimates out-of-sample performance using all the data.
It assumes samples are independent. Overlapping labels or serial dependence, as in most market data, break that assumption, and an ordinary split overstates performance. See purged cross-validation.
Example: 1,000 observations in 5 folds give 5 test sets of 200 observations each.