After a purged fit, the researcher will ask which inputs drove the meta-model. The vault's ML library supports a limited answer: random forests expose impurity-based
feature_importances_, and the bagging lesson warns that ensembles obscure single-tree structure while still ranking predictors. It does not derive López de Prado's MDI/MDA/SFI estimators, clustered importance or substitution-effect algebra. This lesson teaches the caution and audit those sources actually support.Learning objectives. By the end of this note you should be able to:
- state what forest impurity importance measures in the library note;
- explain why correlated or transformed duplicates confuse rankings;
- design an audit checklist for the event classifier without inventing MDA formulae;
- refuse over-confident "driver" claims in the research dossier.
In this lesson
- What the vault actually says
- The decision: audit, do not mythologise rankings
- Deliverable: importance caveat block
Before you start
The full lesson, with worked examples and exercise solutions, is for members.
Read the full lessonNew to SelfMadeQuant.com? See membership options or browse the Financial Machine Learning syllabus.