A systematic researcher building an event classifier for a long–short equity book does not start with a generic sklearn template. The first decisions are financial: which observations count as events, what economic outcome the label measures, how overlapping outcomes share information, and how much memory features retain after stationarity preprocessing. This topic sequences those decisions before any ensemble is fit.
By the end of this note you should be able to:
- choose an event sampling scheme (time versus activity bars) for the classifier's feature clock;
- choose a label that matches the intended exit policy, including first-barrier versus fixed-horizon targets;
- separate primary-side prediction from secondary-size (meta-labeling);
- weight overlapping samples using concurrency diagnostics without confusing weights with leakage control;
- compute fractional-differencing weights, apply a fixed-width truncation, and choose a training-sample order without treating an ADF rejection as a stationarity proof.
In this lesson
- Contents
- The big picture
The full lesson is for members.
Read the full lessonNew to SelfMadeQuant.com? See membership options or browse the Financial Machine Learning syllabus.