Bet sizing consumes a predicted probability of a correct side. Classifier evaluation shows that scores can rank well (high AUC) while a chosen threshold still misbalances precision and recall. The vault does not derive Platt scaling, isotonic regression or reliability-diagram mathematics. This lesson therefore teaches why an uncalibrated score is not a probability input to sizing, and what the researcher must check — without inventing calibration estimators.
Learning objectives. By the end of this note you should be able to:
- separate ranking quality from probability meaning;
- state what bet sizing assumes about ;
- list empirical checks on purged OOS scores before sizing;
- refuse to treat
predict_probaoutput as automatically Kelly-ready.
In this lesson
- Scores that rank are not automatically probabilities
- The decision: gate sizing on empirical reliability
- Deliverable: calibration checklist
Before you start
- Meta-labeling
- Evaluating Classifiers: Precision, Recall, and ROC/AUC
- Bet Sizing: Translating Predicted Probabilities into Positions
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.