Overview

With labels, weights and memory frozen, the systematic researcher still needs an honest estimate of whether the event classifier generalises. Ordinary random $k$-fold assumes IID folds; overlapping financial labels violate that. This topic contrasts purged/embargoed validation with library $k$-fold, then audits feature importance under the limited statements the vault actually supports.

By the end of this note you should be able to:

• draw a purged-and-embargoed split for the running fixture;
• explain why library $k$-fold overstates skill on overlapping labels;
• run an importance audit that respects correlated substitutes without inventing MDI/MDA formulae absent from the vault.

1. Contents
2. The big picture