Overlapping first-barrier labels share return increments. Counting five fixture events as five independent draws inflates the apparent sample size and makes a random train/test split leak: a training label can share the same market move as a test label. This note adapts concurrency and average uniqueness from labels and overlapping outcomes, and links them to purging. It does not invent sequential-bootstrap or uniqueness estimators beyond what those notes give.
Learning objectives. By the end of this note you should be able to:
- compute concurrency and average uniqueness on the fixture;
- use uniqueness as a training weight diagnostic, not as a substitute for purging;
- explain why a random split leaks when labels overlap;
- state the deliverable weight table for the event classifier.
In this lesson
- Concurrency is why random splits fail
- The decision: weight inside train, purge across folds
- Deliverable: weight ledger
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.